Privacy compliance judgment method based on role-based asymmetric adjudication and storage medium

Abstract

The application discloses a privacy compliance judgment method based on role-based asymmetric ruling and a storage medium, and relates to the field of information security.The method is executed by a computing device, and sequentially performs declarative chapter filtering, investigation case generation, role-based small sample retrieval, law evidence retrieval, prosecution doubt generation and support branch control, defense structured refutation and theme-aware refutation branch calculation, asymmetric formalized ruling and judgment reason explanation, and outputs reproducible, traceable and interpretable compliance judgment results. Among them, the role-based retrieval is realized by a text coding neural network, the prosecution, defense and judgment reason are generated by a large language model decoding network, and the formalized ruling is completed by a deterministic rule engine according to a preset threshold, weight and indicator function. The coding network and the decoding network are deployed on a central processing unit or a graphics processing unit to perform forward propagation, and the rule engine performs deterministic calculation on the intermediate data of the network and outputs the final ruling.

Description

Technical Field

[0001] This invention belongs to the field of information security technology, and specifically relates to a privacy compliance judgment method and storage medium based on role-based asymmetric adjudication. Background Technology

[0002] Mobile applications, internet platforms, and data processors typically disclose their personal information processing activities to users through privacy policies, user agreements, children's privacy protection rules, permission request pop-ups, third-party SDK lists, and personal information sharing lists. Whether these texts meet the requirements of legality, legitimacy, necessity, good faith, transparency, clear purpose, minimum necessity, protection of user rights, protection of minors, protection of sensitive personal information, and notification of cross-border provision directly affects the compliance risks of data processing activities.

[0003] Current privacy compliance review methods mainly include manual review, keyword rule review, ordinary machine learning classification, single-model question-and-answer review, and ordinary retrieval-enhanced generative review. Manual review is costly, time-consuming, and difficult to guarantee standard consistency; keyword rule review struggles to handle context, exceptions, chapter positions, and paragraph semantics; ordinary machine learning classification relies on large amounts of labeled data, while privacy compliance labeling is costly, involves detailed violation types, and is subject to rapid legal updates; single-model question-and-answer review is prone to legal illusions, reasoning drift, and unreproducible conclusions; while ordinary retrieval-enhanced generative review incorporates legal provisions, it typically still relies on the large language model to directly output the final conclusion, lacking a formally verifiable adjudication function. In recent years, solutions based on Large Language Models (LLMs) have further developed upon these foundations, but they still suffer from the common problem of conclusions being directly given by the model and difficult to reproduce. Furthermore, they fail to simultaneously address the issues of role confusion and the asymmetry between evidence reliability and error costs in low-resource Chinese compliance review, as detailed below: Regarding privacy policy compliance assessment based on large language models, existing solutions intelligently segment privacy policies to alleviate the problems of long text truncation and forgetting. They also construct assessment knowledge graphs and navigation semantic trees, using vector matching and filtering to achieve structured mapping and tracing from legal rules to specific policy clauses, thereby generating structured assessment records. See patent document CN122021646A for a method, device, medium, and product for privacy policy compliance assessment based on a large model. However, this approach still directly outputs assessment conclusions by comparing the large language model with the knowledge graph and semantic tree, lacking a formal adjudication function that can be independently reproduced by a rule engine. It also fails to introduce a small-sample mechanism for isolating positive and negative examples and a mechanism to address the asymmetric costs of missed and false positives. Another solution reconstructs network traffic from connected vehicle devices and extracts message body data. After initial screening using pattern matching, it performs deep semantic recognition using a large language model with fine-tuned parameters. Then, it compares the privacy transmission behavior in actual traffic with the privacy policy statement to generate a compliance detection report. See patent document CN122069079A for a method for analyzing privacy leaks in connected vehicle devices based on a large language model. This approach relies on actual operational traffic data collection, making it difficult to provide traceable and reproducible compliance rulings for a single paragraph when only policy texts and low-label resources are available.

[0004] In terms of compliance assessment and legal reasoning, existing solutions employ Large Language Models (LLM-as-a-Judge) or Multi-Agent Debate to evaluate texts. For example, existing solutions construct a courtroom-style multi-agent debate framework, setting roles such as plaintiff, defense, and judge to conduct structured adversarial verification of disputed claims, and introducing progressive retrieval to enhance the dynamic expansion of the evidence pool (see arXiv:2603.28488, 2026); existing solutions propose a multi-agent debate framework oriented towards judges and employ an adaptive stopping mechanism (see arXiv:2510.12697); and other solutions propose a cost-sensitive adversarial evaluation framework, introducing role division and bias auditing for defense counsel, judges, and jurors (see arXiv:2410.04663). The above schemes have verified the effectiveness of role-based adversarial and structured adjudication, but their adjudication conclusions are usually still given by a comprehensive large language model, and often adopt majority voting, weighted average or symmetrical debate aggregation, which does not fully reflect the asymmetry that the cost of missed judgment is higher than the cost of false positive in privacy compliance; moreover, its role retrieval does not distinguish the different uses of positive examples of violations and negative examples of compliance, does not combine the chapter review context, and lacks cost-sensitive judgment thresholds corresponding to the costs of missed judgment and false positive.

[0005] Regarding few-shot and retrieval enhancement, existing solutions dynamically select examples through vector similarity retrieval to improve the learning effect of few-shot context. However, most of these solutions use a single similarity retrieval method, failing to distinguish the roles and uses of positive and negative examples, or consider the context of chapter review. This can easily lead to the mixing of non-compliant and compliant samples, role confusion, and cross-chapter misleading in compliance reviews. Early research also attempted to achieve automatic detection of privacy risks from the perspectives of detecting internal contradictions in privacy policies and comparing policy statements with application code or data streams, but these studies mostly focused on code behavior analysis or rule matching.

[0006] In privacy compliance assessments, determining whether a paragraph violates regulations is not a simple binary classification. The system first needs to secure the original text evidence and identify the data processing scenarios involved. Secondly, it needs to raise potential issues of illegality or insufficient disclosure. Thirdly, it needs to match these issues with legal basis, regulatory norms, violation patterns, or historical samples. Next, it needs to determine whether the paragraph contains facts sufficient to refute the issues, such as functional necessity, clarity of authorization and purpose, user control, and the principles of minimum necessity and proportionality. Only then should a conclusion be reached. Skipping these intermediate steps and directly outputting "compliant" or "non-compliant" by the model makes the process difficult to audit and hinders subsequent manual review and experimental evaluation.

[0007] Meanwhile, privacy compliance assessments suffer from an asymmetric cost of error. Misjudging genuinely non-compliant clauses as compliant may lead to infringement of personal information rights, regulatory penalties, rectification costs, reputational damage, and subsequent litigation risks; misjudging genuinely compliant clauses as risky typically incurs costs primarily for review and text optimization. Therefore, the cost of omitting a violation in compliance assessments usually outweighs the cost of misreporting. Existing methods such as symmetrical debate, majority voting, or simple average scoring tend to assign equal weight to the prosecution's doubts and the defense's arguments, failing to fully reflect this asymmetric cost of error.

[0008] Furthermore, in low-resource compliance review scenarios, few-shot learning cannot be simply understood as adding a number of random cases to the prompts. The prosecution needs similar violation samples to help identify inconsistencies, while the defense needs similar compliance samples to help identify refutable facts. Mixing violation and compliance samples can easily lead to role confusion and a shift in judgment boundaries. Moreover, different sections of the privacy policy have different review contexts; for example, the compliance elements differ in sections such as "Rights of Personal Information Subjects," "Protection of Minors," "Third-Party Sharing," and "Storage and Protection." Ignoring the section context can easily result in retrieving semantically similar samples with different review tasks, leading to false positives or false negatives.

[0009] The final compliance conclusions of the aforementioned existing solutions are mostly generated directly by large language models. Their outputs fluctuate with the prompt words and sampling process, making it difficult to reproduce and formally verify at the computational level. Furthermore, it is impossible to distinguish the different costs of missed judgments and false alarms in the adjudication calculation. This constitutes a technical defect of the existing technology in terms of the reproducibility and verifiability of the conclusions. Summary of the Invention

[0010] In existing privacy compliance reviews based on large language models, the final compliance conclusion is usually directly generated by the large language model. This is subject to technical defects such as unreproducible conclusions, lack of formal verification, unstable evidence sources, and the inability to differentiate and control the costs of missed judgments and false alarms at the computational level, due to the influence of prompt words, sampling randomness, and model discretion. The purpose of this invention is to provide a privacy compliance judgment method based on role-based asymmetric adjudication, and a corresponding computer-readable storage medium. This method is executed by a computing device. After inputting the text to be reviewed, the computing device's processor sequentially executes the following steps: declarative chapter filtering, investigation file generation, role-based retrieval of small samples of violations and compliance, legal evidence retrieval, generation and support gate control of prosecution doubts, calculation of structured rebuttal and topic-aware rebuttal of the defense, asymmetric formal adjudication, and explanation of the judgment's reasoning. This results in a reproducible, traceable, and interpretable compliance judgment. The role-based retrieval of small samples in the categories of violations and compliance is implemented by a text encoding neural network. The generation of prosecution doubts, the structured rebuttal of the defense, and the explanation of the judgment's reasons are implemented by a large language model decoding network. The asymmetric formal adjudication is implemented by a deterministic rule engine running on a processor. The aforementioned text encoding neural network and large language model decoding network are deployed on the central processing unit or graphics processing unit of the computing device to perform forward propagation. The rule engine performs deterministic calculations on the intermediate data output by the network based on preset thresholds, weights, and indicator functions, and then outputs the final adjudication result.

[0011] The objective of this invention is achieved through the following technical solution: According to a first aspect of the present invention, a privacy compliance judgment method based on role-based asymmetric adjudication is provided. The method is executed by a computing device, which includes a processor, a text input interface to be reviewed, a legal evidence library, a small sample case library, a vector retrieval module, and a large language model interface. The method includes: Acquisition steps: Acquire the text to be reviewed through the text input interface, and preprocess the text to be reviewed to extract at least one unit to be reviewed; Search steps: Based on the unit to be reviewed, semantic similarity search is performed in the legal evidence database and the small sample case database through the vector retrieval module to obtain the legal provision-related evidence set and case-related evidence set corresponding to the unit to be reviewed; Integration Step: Input the unit to be reviewed, the legal provision-related evidence set, and the case-related evidence set into the large language model interface, and generate a ruling result that integrates the legal provision-related evidence set and the case-related evidence set through the large language model; The method further includes: The text to be reviewed is segmented into units and its metadata is annotated to obtain at least one unit to be reviewed; For each of the aforementioned entities to be examined, a role-based adversarial process and asymmetric adjudication process are executed, the process including: The sub-step for generating investigation files is as follows: Generate case file data for the current unit to be reviewed, including the processing scenario, fragments of original evidence, a set of inspection results, clues to doubts, facts favorable to the defense, and questions to be asked by the prosecution. Role-based retrieval sub-step: Retrieve violation-type small samples and compliance-type small samples from the small sample case library respectively; Evidence aggregation sub-step: Retrieve legal evidence from the legal evidence database and merge it with the sample-labeled evidence extracted from the small sample of violations to form a legal evidence set; The prosecution's sub-step involves inputting the current unit to be reviewed, the small sample of violations, the legal evidence set, and the case file data into the prosecution's intelligent agent to generate at least one structured point of doubt. The prosecution's intelligent agent is implemented by a large language model decoding network, and the legal basis it cites must belong to the legal evidence set. The sub-step for determining the validity of doubts is as follows: a support score is calculated for each structured doubt. The support score is used to characterize the comprehensive support of the doubt in three dimensions: legal basis, violation cases, and text gaps. The valid doubts are determined based on the comparison result between the support score and the preset support threshold. The defense's rebuttal sub-step involves inputting the valid doubts, compliance-related small samples, legal evidence sets, and case file data into the defense agent to generate a rebuttal response for each valid doubt. The rebuttal response includes at least the doubt number, the subject of review, the defense arguments, the original text, the cited legal provision number, the cited compliance sample number, the uncovered prosecution points, and four quantitative evaluation items. These four quantitative evaluation items are: functional necessity evaluation item, authorization and purpose clarity evaluation item, user control evaluation item, and minimum necessity and proportionality principle evaluation item, respectively used to characterize the degree of compliance of the unit under review in terms of functional necessity, authorization clarity, user control, and proportionality principle. The sub-step for determining the validity of a rebuttal is as follows: Based on the review topic to which the current unit to be reviewed belongs, assign corresponding topic weights to the four quantitative evaluation items, calculate the rebuttal score by weighting, and determine whether the doubt has been successfully refuted based on the comparison result between the rebuttal score and the preset rebuttal threshold. Asymmetric adjudication sub-step: By using the rule engine running on the processor, valid doubts that were not successfully refuted are counted. If there is at least one valid doubt that was not successfully refuted, a non-compliant conclusion is output; otherwise, a compliant conclusion is output. The adjudication result is not directly generated by the large language model. The output step involves inputting the ruling result, valid doubts, defense responses, unresolved doubts, and rebuttal points into the adjudicator's intelligent agent to generate the adjudication reasoning in natural language.

[0012] Furthermore, the sub-step for generating the investigation case file also includes: Based on the chapter title or paragraph topic of the unit to be reviewed, determine whether it belongs to a declarative or administrative chapter; If so, skip the subsequent prosecution and defense process, directly output a low-risk compliance conclusion, and record the reason for skipping in the review log; If not, then the case file data is generated.

[0013] Furthermore, in the role-based retrieval sub-step: The violation-related small samples are retrieved only from the sample pool marked as non-compliant in the small sample case library; The compliance-related small samples are retrieved only from the sample pool marked as compliant in the small sample case library; Furthermore, when performing the comparison document screening, samples with the same chapter title or the same review scenario as the current unit under review are selected as the first comparison benchmark; interference samples with the same policy source identifier as the current unit under review are excluded from the first comparison benchmark, and the target comparison document set is constructed from the remaining samples other than the interference samples.

[0014] Furthermore, the support score is calculated in the following manner: For each structured point of doubt, calculate its legal provision matching degree with the legal evidence set, its case similarity with the small sample of violations, and its text missing intensity in the current unit to be reviewed; The support score is obtained by weighted summation of the legal provision matching degree, case similarity, and text missing intensity; the weights of the legal provision matching degree, case similarity, and text missing intensity are all preset non-negative values, and the sum of the three is 1; The calculation method for the legal provision matching degree is as follows: Obtain the legal basis number cited for the structured doubts; Search for the legal citation scores corresponding to each legal basis number in the legal evidence set, take the highest score and normalize it, and use the normalized value as the legal citation matching degree; If the legal basis number cited in the structured doubt is not in the set of legal evidence, then the legal provision matching degree is 0.

[0015] Furthermore, the rebuttal score is calculated as follows: Based on the review topic to which the current unit to be reviewed belongs, determine the four topic weights that correspond one-to-one with the four quantitative evaluation items; The rebuttal score is obtained by multiplying the value of each quantitative evaluation item by its corresponding topic weight and then summing the results. The values ​​of the four quantitative evaluation items are all selected from a set of six preset anchor points. The input prompts of the defense agent contain preset scoring specifications for each evaluation dimension and six anchor point judgment criteria. The defense agent outputs the values ​​of each quantitative evaluation item, the span of the original text evidence on which it is based, and the reasons for the score according to the specifications, and writes them into the structured review log.

[0016] Furthermore, in the defense's rebuttal sub-step, evidentiary constraints are applied to the defense response: If no corresponding original evidence span is given for any quantitative evaluation dimension, the score for that dimension will be capped at the first preset upper limit. If the defense response does not cite any legal evidence or any compliance-related small samples, then the rebuttal score will be capped at the second preset upper limit; If the defense response lists prosecution points that are not covered, the rebuttal score will be deducted according to a preset deduction coefficient.

[0017] Furthermore, in the asymmetric decision sub-step: The rule engine reads the semantic similarity data output by the vector retrieval module, the structured doubt data output by the prosecution agent, and the structured defense response data output by the defense agent. The rule engine outputs the final ruling result according to the structured question data and the structured defense response data, using deterministic calculation logic, so that the same input data, calculated by the rule engine under the same parameters, yields the same ruling result.

[0018] Furthermore, the method also includes a threshold dynamic self-correction step: The computing device maintains a verification buffer for storing historical review records with manual review labels; When the number of labeled records in the verification buffer reaches a preset lower limit, multiple sets of candidate support thresholds and candidate rebuttal thresholds are traversed in the candidate threshold grid, and the empirical asymmetric risk under each set of candidate thresholds is calculated respectively. Select a set of candidate thresholds that minimizes the empirical asymmetric risk as the optimal threshold for this round; The current threshold is updated using an exponential moving average with a preset learning rate, and the updated threshold is then clipped to a preset range.

[0019] Furthermore, the prosecution's intelligent agent, the defense's intelligent agent, and the judge's intelligent agent are each implemented by a large language model decoding network. The large language model decoding network is a Transformer neural network with a decoder architecture, which generates candidate outputs in an autoregressive manner. The fields and formats of the candidate outputs are limited by preset output rules in the input prompts, and the processor performs structural parsing and field verification to form structured data that conforms to the preset format. The vector retrieval module is implemented by a text encoding neural network with a Transformer encoder architecture, which is used to map text fragments into dense semantic vectors.

[0020] According to a second aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, which, when executed by a processor, cause the processor to perform the steps of the privacy compliance judgment method based on role-based asymmetric adjudication.

[0021] Through the above design scheme, the beneficial effects of this invention compared with the prior art are as follows: First, this invention decomposes privacy compliance judgment into independently verifiable stages such as evidence fixation, doubt generation, evidence gating, adversarial rebuttal, and formal adjudication. The criteria and intermediate results of each step are explicitly recorded, so that the final conclusion can be gradually traced back to specific evidence and doubts. This overcomes the problems of invisible intermediate reasoning and unverifiable conclusions in single-model question-and-answer review, and improves traceability and auditability. Second, this invention uses declarative section filtering to avoid excessive nitpicking of paragraphs that do not describe specific data processing activities, such as basic declarations or organizational information, thereby reducing meaningless false alarms. Third, this invention uses chapter-aware small sample retrieval for violations and compliance to enable violation samples and compliance samples to be used for the discovery of doubts and the restoration of refutations, respectively, and ensures that the context of the samples is consistent with the chapter theme of the paragraph to be reviewed, thereby reducing role confusion and cross-chapter misleading in ordinary small sample methods. Fourth, this invention reduces the illusion of legal provisions by merging manually annotated legal provisions on the samples of violations with the results of legal provision RAG (retrieval-enhanced generation) and restricting the prosecution to citing only legal provisions in the evidence set; Fifth, this invention, through the formal calculation of support scores and rebuttal scores, and the asymmetric adjudication function executed by the rule engine, enables the validity of doubts, their rebuttal, and the final adjudication result to be reproduced and verified on the computing device by thresholds, weights, and evidence constraints. Compared with the solution that directly outputs compliance conclusions from large language models, this invention reduces the technical defect of unreproducible conclusions and improves the stability, verifiability, and auditability of the judgment results. Sixth, this invention uses four quantitative evaluation items based on topic perception. N / A / U / P Weighting allows for different approaches to rebuttal on different review topics. For example, user rights topics receive more emphasis. U Third-party shared and storage themes place greater emphasis on A The theme of security measures is more emphasized. N and A This improves the adaptability of the rebuttal points; Seventh, this invention limits the defense from giving high scores when there is no original evidence, legal evidence or compliant sample to support it, by using six-level scoring anchor points, capping the span of dimensional evidence, capping the lack of valid citations, and deducting points that are not covered. This reduces the chance of missed judgments. Eighth, by using an asymmetric adjudication function, any valid and unrefuted point of contention can trigger a non-compliance conclusion, reflecting the asymmetric cost of missed judgment being higher than the cost of false positive. This allows the present invention to control the false positive rate while maintaining a high violation identification capability. Compared with symmetric aggregation schemes such as symmetric debate, majority voting, or simple averaging, it achieves a better balance in comprehensive indicators (such as F1 score and false positive rate). Ninth, this invention supports the recording of scores, rebuttal scores, legal evidence numbers, small sample numbers, unresolved doubts, running time, model usage, and judgment reasons through structured log records, which facilitates auditing, review, experimental evaluation, and regulatory record keeping. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of this application. The illustrative embodiments and descriptions of the invention are used to understand the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is an overall architecture diagram of the privacy compliance judgment system in this embodiment of the invention; Figure 2 This is a flowchart illustrating the overall execution of the privacy compliance judgment method in this embodiment of the invention. Figure 3 This is a flowchart of the investigation case file generation process in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the merging of role-based small samples and legal evidence in an embodiment of the present invention; Figure 5This is a flowchart illustrating the generation of prosecution doubts, filtering of legal provisions numbers, and support for segmented gate control in this embodiment of the invention. Figure 6 This is a flowchart illustrating the compliance protection process in an embodiment of the present invention; Figure 7 This is a flowchart of the scoring process for the defense's arguments and rebuttals in an embodiment of the present invention; Figure 8 This is an asymmetric formalized adjudication flowchart in an embodiment of the present invention; Figure 9 This is a structured review log field relationship diagram in an embodiment of the present invention; Figure 10 This is a comparison chart of the method of the present invention with random retrieval, similarity retrieval, use of only legal evidence, symmetric debate, and ablation configuration on the main evaluation indicators; Figure 11 This is a graph showing the changes in recall rate, false positive rate, F1 score, and total cost-sensitive risk as a function of the rebuttal threshold in this embodiment of the invention. Figure 12 This is the model network structure and data flow diagram in the embodiments of the present invention; Figure 13 This is a privacy compliance judgment state diagram topology diagram in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] like Figures 1 to 13 As shown, this invention proposes a privacy compliance judgment method based on role-based asymmetric adjudication. The method is executed by a computing device, which includes a processor, a text input interface to be reviewed, a legal evidence library, a small sample case library, a vector retrieval module, and a large language model interface. The method includes: Step 1: Obtain the text to be reviewed, and segment it according to paragraphs, clauses, heading levels, or semantic boundaries to obtain the units to be reviewed. For each unit to be reviewed Generate basic metadata, which includes at least one of policy identifier, paragraph number, chapter title, and paragraph text.

[0025] Step 2: Determine the unit to be reviewed based on the chapter title or paragraph topic. Whether it belongs to a declarative section or an administrative section; if it is a unit to be reviewed If the passage is a declarative or administrative section, skip the subsequent prosecution and defense process, directly output a low-risk compliance conclusion, and record the reason for skipping in the review log; if it is not a passage in the above categories, proceed to step 3.

[0026] Step 3: Unit to be reviewed The execution of the investigation case file generation, obtaining the case file ;files This includes processing scenario sets, original evidence fragment sets, inspection result sets, suspicious clue sets, sets of facts favorable to the defense, and sets of questions to be asked by the prosecution; this step only fixes the textual facts and review clues, and does not output illegal conclusions, compliance conclusions, or risk levels.

[0027] Step 4: Based on the unit to be reviewed ,files Retrieve violation-related small samples from a small sample case library, including chapter titles and section titles. and compliance-related small samples ;in, The samples are retrieved only from a pool marked as non-compliant, for use by the prosecution in generating allegations. The samples are retrieved only from the pool marked as compliant, for use by the defense in its arguments and rebuttals; samples with the same chapter titles or the same review scenarios are given priority.

[0028] Step 5: Retrieve legal evidence from the legal evidence database And from a small sample of violations Sample annotation evidence was extracted from manually annotated legal provisions fields. ,Will and By removing duplicates and merging them according to their evidence numbers, we obtain the legal evidence set. Specifically, during the merging process, the sample labels are first used to obtain evidence. Add the candidate evidence to the set in the order of its generation, and then add the legal provision retrieval evidence. Add evidence in order of search ranking; during the addition process, use the evidence number as the unique key for deduplication. and If there are identical evidence numbers, then retain them. The sample evidence is labeled and the corresponding legal provision retrieval evidence is discarded; when the number of evidences in the candidate evidence set reaches the preset upper limit... Stop adding subsequent evidence at that time, among which Set it to 5.

[0029] Step 6: Select the unit to be reviewed Small sample size for violations Collection of legal evidence and case files Input the controlling agent to generate one or more structured questions. Each structured point of doubt The prosecution's AI agent must include at least the number of the alleged violation, the type of violation, the alleged claim, the number of the legal basis, and a fragment of the original evidence. The final compliance conclusion, risk level, or rectification conclusion must not be output by the prosecution's AI agent. The legal basis number cited by the prosecution must belong to the legal evidence set. It does not belong to The legal provision numbers are filtered; the controlling agent is implemented by a large language model decoding network. The processor, according to a preset prompt template, processes the unit to be reviewed. Small sample size for violations ,files and collection of legal evidence The input lexical sequence is concatenated and encoded. The preset prompt template specifies at least the prosecution role, field names and types of structured doubts, generation requirements for each field, format instructions to output only data in a preset structured format, and prohibited output content. The large language model decoding network performs forward computation on the input lexical sequence and generates candidate outputs in an autoregressive manner, using existing large language model structured output technology. The processor performs structure parsing and field verification on the candidate outputs; when a candidate output conforms to the preset structured format and contains a doubt number, violation type, doubt claim, legal basis number, and original evidence fragment, it is identified as a structured doubt. For those not belonging to the collection of legal evidence The legal basis number is used for filtering. For candidate outputs that cannot be parsed into the preset structure or lack required fields, format repair, regeneration, or discarding can be performed. The above-mentioned structured output method, which limits the output format based on prompt templates and parses and verifies the generated results, is implemented using existing large language model structured output technology in this field.

[0030] Step 7: For each structured point of doubt Calculate support score And based on support for Determine whether the point of contention is valid; support the analysis. satisfy: (1); in, This indicates that the prosecution's intelligent agent is targeting the unit to be reviewed. A structured point of contention was generated. This represents a unit in the text to be reviewed. Indicating structured doubts The support score has a value range of [0,1] and has no physical unit; Represents a collection of legal evidence abbreviation; Indicating a small sample of violations abbreviation; Indicating structured doubts With legal evidence collection The legal provision matching degree has a value range of [0,1] and has no physical unit; Indicating structured doubts Small sample size of violations The similarity or type consistency is a value ranging from [0,1] and has no physical unit. Indicating structured doubts In the unit to be reviewed The strength of the corresponding text gap, with a value range of [0,1], has no physical unit; , , This represents the preset weight, with values ​​ranging from [0,1]. Equation (1) addresses the issue of unstable evidence sources when a single model directly provides a conclusion by binding three technical features—legal provision matching, similarity of small samples of violations, and gaps in the original text—to the text processing workflow under review; when Greater than or equal to the support threshold At that time, determine If the point of contention is valid, then it is considered valid; otherwise, it is determined to be valid. This point of contention will not be included in the subsequent ruling, that is... This function indicates the validity of the doubtful point and takes the value of 0 or 1.

[0031] Specifically, the degree of legal provision matching Calculate as follows: For structured doubts Each legal basis number cited is part of the legal evidence collection. The search results for the legal provision corresponding to the given number are retrieved, and the highest score is taken, normalized, and then restricted to the interval [0,1] as the maximum score. The retrieval score for legal provisions is obtained by weighting keyword matching degree, number of check item hits, and clause importance; if structured doubts... The cited legal provisions are not listed. In the middle, then Small sample similarity of violation categories In one feasible manner, the unit to be reviewed is taken. Small sample size of violations The maximum cosine similarity of the text samples; in another possible approach, the structured suspicious points are taken. The maximum similarity between the corresponding text fragment and each violation sample is used to achieve suspicious-level matching; text gap strength. Used to measure structured doubts The elements to be disclosed are in the unit under review. The degree of missing information, based on structured doubts The corresponding inspection items are defined as the proportion of elements not disclosed in the original document among the elements that should be disclosed (such as the purpose of processing, recipient, type of personal information, retention period, authorization method, or consequences of refusal) to the total number of elements that should be disclosed. The value range is [0,1]; it takes 1 when all elements to be disclosed are missing, and 0 when all have been disclosed; in a simplified implementation, it takes a higher value when the structured doubt provides evidence pointing to the gap in the original text, and a lower value when it does not.

[0032] Step 8: Implement false positive suppression rules for the prosecution's suspicions; if the suspicion falls under a preset narrow-range false positive pattern, for example, the narrow-range false positive pattern includes the following situations: First, in personal information transfer scenarios such as mergers, acquisitions, or bankruptcy liquidation, the future recipient has not yet been determined, but the text under review has clearly promised that the recipient will continue to be bound by the original privacy policy and will obtain the individual's consent again when the recipient changes the original processing purpose or processing method; in this case, the suspicion is not established simply because the text under review does not list the undetermined recipient in advance; Second, the text under review lists situations where personal consent is not required as stipulated by laws and regulations. The prosecution's concerns are solely based on the fact that the text fails to explain the specific purpose, method, or exercise of individual rights under each statutory circumstance, and the text under review does not add any circumstances exceeding the scope of statutory exceptions. In this case, the concerns are not established simply because the text under review does not further elaborate on the statutory exceptions. Third, the text under review has explained the criteria for determining the retention period and promised to delete, anonymize, or de-identify personal information after the retention period expires. The prosecution's concerns are solely based on the fact that the text does not provide a uniform and fixed number of retention days. In this case, the concerns are not established simply because the retention period is not expressed in a specific number of days.

[0033] Step 9: Select the unit to be reviewed Valid set of doubts, small sample of compliance issues Collection of legal evidence and case files Input the defense agent and generate a rebuttal response for each valid point of contention. ; defense response It should include at least the number of the point of contention, the subject of the review, the defense arguments, the original text, the number of the cited legal provision, the number of the cited compliant sample, the key points of the prosecution that were not covered, and four quantitative evaluation items. , , , ;in Indicates functional necessity. This indicates that the authorization and purpose are clear. Indicates user control. This represents the principle of minimum necessity and proportion; each evaluation item preferably takes a value from a six-level anchor point set {0, 0.2, 0.4, 0.6, 0.8, 1}; the defense agent is implemented by a large language model decoding network, which is a Transformer neural network with an existing decoder architecture, and incorporates effective doubts and compliance-related small samples. The legal evidence is encoded as an input word sequence, and after forward propagation, the above-mentioned defense response is output under structured constraint decoding. It includes scores for its four quantitative evaluation dimensions.

[0034] Step 10: Response to each defense Calculate topic perception rebuttal score First, based on the review topic Obtain weight Then calculate: (2); In equation (2), This represents the structured defense response generated by the defense agent in response to a valid point of contention. Indicate a defense response The rebuttal score has a value range of [0,1] and has no physical unit; The subject of the review is indicated, with values ​​including collection and use, storage, security measures, third-party sharing, user rights, protection of minors, cross-border transfer, or sensitive personal information. Indicates the subject of review The corresponding weight vector; , , , These represent the weights of four quantitative evaluation items—functional necessity, authorization and purpose clarity, user control, and minimum necessity and proportionality principle—under the theme, with values ​​ranging from [0,1]. ; , , , The scores represent the functional necessity, authorization and purpose clarity, user control, and minimum necessity and proportionality principle, respectively, with a value set of {0, 0.2, 0.4, 0.6, 0.8, 1}, and no physical units; Equation (2) limits the rebuttal score to the calculation process constrained by the evidence span, compliance samples, and review topics, in order to suppress high-scoring defenses without evidence; if the review topic is missing or not in the preset topic set, the default weight is used; specifically, the input prompts of the defense agent preset the scoring specifications for each evaluation dimension and the six-level anchor point judgment criteria, and the scores are based on these specifications. , , , Output the values ​​of the six anchor points respectively, and write the score of each dimension, the span of the original evidence on which it is based, and the reason for the score into the structured review log for review and reproduction.

[0035] Step 11: When calculating the rebuttal score, apply evidence constraints; if a certain evaluation dimension does not provide a corresponding original text evidence span, then the score of that evaluation dimension is limited to no higher than a first preset upper limit; in one embodiment, the first preset upper limit is 0.4; calculate the rebuttal score based on the scores of each evaluation dimension after constraints and their corresponding weights. If the defense response does not cite any legal evidence and does not cite any compliance-related small samples, then the rebuttal score is limited to no higher than a second preset upper limit; in one embodiment, the second preset upper limit is 0.6; the capping process only truncates scores exceeding the corresponding preset upper limit, and does not increase scores below the corresponding preset upper limit; after the above capping process, the rebuttal score is limited to the [0,1] interval.

[0036] Step 12: When Greater than the rebuttal threshold At that time, determine Structured doubts Successfully refuted; otherwise confirmed. If a valid point of contention does not have a corresponding defense response, it is considered... .

[0037] Step 13: Perform asymmetric adjudication; let the set of valid and unrefuted points of contention be: (3); In equation (3), Indicates the unit to be reviewed A set of valid, unrefuted doubts, the elements of which are structured doubts. It has no physical unit; This function indicates the validity of the doubtful point and takes a value of 0 or 1. This is a function indicating that the doubt has been successfully refuted; its value is either 0 or 1. Representation and Structured Questions Corresponding defense response; if If not empty, output an invalid conclusion; if If empty, output a compliance conclusion; that is: (4); In equation (4), This represents the asymmetric decision function for the unit under review, with a value of 0 or 1, where 1 indicates non-compliance and 0 indicates compliance. This indicates an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 if it is false. This indicates the existence of at least one structured suspicious point. ; and The meaning is consistent with Equation (3); the ruling result is calculated by the code or rule engine running on the processor, and is not directly generated by the large language model. Therefore, Equation (4) is not a pure mathematical rule detached from the application scenario, but a data processing step bound to the privacy compliance text, legal evidence set, prosecution doubts and defense response; the ruling function is output by the code or rule engine on the processor after calculating the values ​​of the valid doubt set and the rebuttal indication function. It does not rely on the discretion of the large language model, thereby overcoming the technical defect that the conclusion cannot be reproduced due to the randomness of sampling and the sensitivity of prompt words when the large language model directly outputs the compliance conclusion, so that the same input can obtain the same ruling result under the same threshold and weight.

[0038] Step 14: Generate a risk level based on the ruling, the number of valid doubts, the number of unresolved doubts, and the type of doubts; preferably, the risk level is low when there are no valid doubts; the risk level is medium or low when there are valid doubts but all of them are successfully refuted; the risk level is medium when there is one unresolved valid doubt; and the risk level is high when there are two or more unresolved valid doubts or when high-risk scenarios such as sensitive personal information, minors, or cross-border transmission are involved.

[0039] Step 15: Input the ruling result, valid doubts, defense response, unresolved doubts, and rebuttal points into the adjudicator's intelligent agent to generate the adjudication reasoning in natural language form; the adjudicator's intelligent agent is only used to interpret the already determined ruling result and shall not modify the ruling result output in Step 13; the adjudicator's intelligent agent is implemented by a large language model decoding network, which is a Transformer neural network with an existing decoder architecture, and only generates natural language adjudication reasoning in an autoregressive manner based on the already determined ruling result, without participating in the calculation of the ruling result.

[0040] Step 16: Write the unit to be reviewed, case file, retrieved violation sample number, compliance sample number, legal evidence number, list of doubts, support score, rebuttal score, unresolved doubts, ruling result, risk level, reasoning for the ruling, running time, model usage, and configuration threshold into the structured review log; the structured review log is used for manual review, statistical analysis, threshold adjustment, experimental evaluation, model auditing, and record keeping.

[0041] Step 17: Perform dynamic self-correction of the threshold; support threshold... and rebuttal threshold The parameters remain unchanged during a single privacy compliance assessment, but are not permanently fixed; the computing device, according to a preset time period, or when the number of newly added review records with manual review labels reaches a preset number, uses structured review logs to support thresholds. and rebuttal threshold Update.

[0042] Computing device maintenance verification buffer And read the most recent data from the structured review log. Each review record includes at least the unit to be reviewed, the supporting score for the doubts, the rebuttal score for the defense response, the system's decision, and the manual review label. , This indicates that the manual review result is non-compliant. This indicates that the manual review result is compliant.

[0043] Pre-set candidate support threshold set and candidate rebuttal threshold set The Cartesian product of the two is then used as the candidate threshold grid. ,Right now Candidate support threshold set and candidate rebuttal threshold set Both are finite sets of values ​​within the interval [0,1]; in one embodiment, both candidate threshold sets are set at an interval of 0.05.

[0044] For candidate threshold grid Each set of candidate thresholds The computing device does not re-call the prosecution and defense agents, but instead reads the support scores, rebuttal scores, and their corresponding relationships already stored in the verification buffer, and recalculates the system adjudication result for each review record according to the asymmetric adjudication rules described in step 13 and equation (18); specifically, when at least one support score is not lower than the candidate support threshold, Furthermore, its corresponding rebuttal score is not higher than the candidate rebuttal threshold. If there are any doubts, the review record will be re-determined as non-compliant; otherwise, it will be re-determined as compliant.

[0045] Based on the recalculated system decision results and the manual review labels, the empirical asymmetric risk corresponding to each group of candidate thresholds is calculated according to formula (21); the empirical asymmetric risk is respectively expressed as the cost of missed judgment. and false alarm costs The records of missed detections and false alarms are weighted, among which... Greater than Select the set of candidate thresholds that minimizes the empirical asymmetric risk as the optimal support threshold and optimal rebuttal threshold for this round; if multiple sets of candidate thresholds correspond to the same empirical asymmetric risk, then the set of candidate thresholds that is closest to the current threshold for this round is selected first; if the distance is still the same, then one of them is selected according to the preset traversal order in the candidate threshold grid.

[0046] The computing device, according to equation (22), uses a learning rate The current support threshold, current rebuttal threshold, and optimal support threshold and optimal rebuttal threshold of this round are updated by performing exponential moving averages to obtain the support threshold and rebuttal threshold to be used in the next round. The updated support threshold is then used to update the support threshold. and rebuttal threshold Each threshold is cut to its own preset value range; the updated threshold is written into the configuration parameters and structured review log for use in the next round of privacy compliance judgment; when the next round of privacy compliance judgment begins, the rule engine reads the updated support threshold and rebuttal threshold from the configuration parameters, which are used for the validity determination of the doubts in equation (16) and the success determination of the rebuttal in equation (17), respectively; the threshold written into the structured review log is used to record the threshold version and audit traces.

[0047] When the number of review records with the manual review label in the verification buffer falls below a preset lower limit At this time, the threshold reassessment and exponential moving average update of this round are not performed, and the support threshold and rebuttal threshold of the previous round are continued to be used. The above empirical asymmetric risk calculation and exponential moving average update are formally defined by equations (21) and (22) in the following text.

[0048] The following formulas are all executed by processors deployed on computing devices. The inputs are the text to be reviewed, legal evidence, sample cases, prosecution's doubts and defense's responses. The outputs are the review logs, the rulings and the reasoning behind the rulings. Each formula is used to define specific data processing steps and is not intended to be protected as a purely mathematical method outside of a privacy compliance review scenario.

[0049] Define the text space to be reviewed as The compliance label set is ,in This indicates that the information is not compliant with regulations. This indicates authenticity and compliance. The small sample case library is defined as... The legal evidence database is The compliance review system is then represented as: (5); In equation (5), This represents the compliance judgment function executed by the computing device; This represents a set of units to be reviewed, where the elements are text fragments and have no physical units. This represents a small sample case library, where the elements are manually labeled samples. This represents a legal evidence database, with elements including laws and regulations, normative documents, regulatory rules, or manually maintained violation patterns. This represents the set of output tags, with values ​​fixed at {0,1}; for any unit to be reviewed System output: (6); In equation (6), This represents a unit to be reviewed; Indicates the system output label, This indicates that the system has determined the behavior to be non-compliant. This indicates that the system has determined the system to be compliant. and The meaning is consistent with that of equation (5).

[0050] Define a small sample case library Each sample is: (7); In equation (7), Indicates the sample sequence number, with a value range of 100. , Represents a small sample case library The number of samples in the data is expressed in units of records. Indicates the first Sample; This represents sample text; Indicates compliance label; Indicates the type of violation or compliance, with values ​​taken from a preset set of types. One of the elements; This represents the set of manually annotated legal provision numbers; The text should include the expert's reasoning. This represents a chapter title or review scene tag, with values ​​from a set of preset scenes. One of the elements; except Except for the unit of the item, the other variables have no physical unit; thus, the violation sample pool and the compliance sample pool are defined: (8); (9); In equation (8), This indicates a small sample pool for violations. A subset of; and The meaning is consistent with equation (7); in equation (9), This indicates a small sample pool for compliance purposes. A subset of; and The meaning is consistent with equation (7); the unit to be reviewed When performing role-based few-sample retrieval, the text vector representation function is defined as follows: The similarity function is: (10); In equation (10), Indicates the unit to be reviewed With sample text The cosine similarity of the two values ​​ranges from [-1, 1] and has no physical unit. This represents the text vector representation function generated by the vector retrieval module. Specifically, for any input text... ,in Unit to be reviewed or sample text First, the input text is segmented by a word segmenter. The process involves converting the sequence into a lexical sequence containing initial aggregated lexical units; inputting this lexical sequence into a Transformer encoder, which then performs lexical embedding, positional encoding, and... The forward propagation of the bidirectional multi-head self-attention layer and the feedforward network layer yields the hidden state sequence of the last layer. The hidden state vectors corresponding to the sentence-first aggregated words in the hidden state sequence are extracted and L2 normalized to obtain the input text. Vector representation of: (11); in, Indicates input text After the first After layer Transformer encoding, the hidden state vector corresponding to the sentence-initial aggregate word; Indicates the number of layers in the Transformer coding layer; The dimension of the hidden state vector; ,and ; In one embodiment, the text encoding neural network employs a BAAI / bge-m3 text encoding network with XLM-RoBERTa as its backbone, using sentence-initial aggregation term pooling, and outputs dimensions... It is 1024; because and After L2 normalization, equation (10) is equivalent to: (12); This embodiment directly uses cosine similarity with a value range of [-1, 1]; in other embodiments, if it is necessary to map it to the [0, 1] interval, a different approach is used. Perform a linear transformation.

[0051] The small sample size for violations is: (13); The small sample size for compliance is: (14); In equation (13), Indicated as unit to be reviewed The retrieved small sample of violations; This indicates sorting by similarity and taking the top results. Item retrieval operators; Indicates the number of small samples of violations used by the prosecution, with a positive integer value and a unit of count; , , and The meaning is consistent with equations (7) to (10); Indicates the unit to be reviewed Chapter titles, review topics, or scene tags; Indicates sample The identifier of the relevant policy; Indicates the unit to be reviewed The identifier of the relevant policy; This indicates that samples from the same policy source are excluded; in equation (14), Indicated as unit to be reviewed A small sample of compliance-related items obtained from the search; This indicates sorting by similarity and taking the top results. Item retrieval operators; This indicates the number of compliance-related small samples used by the defense, and the value is a positive integer, in units of samples. The meaning of is consistent with equation (9); the meanings of the remaining variables are consistent with equation (13). In equations (13) and (14) This indicates that the sample and the unit to be reviewed are identical in the normalized chapter title, review topic, or scene label.

[0052] Define the legal evidence set as: (15); In equation (15), Indicates the unit to be reviewed The corresponding set of legal evidence; This represents the maximum number of legal evidence items to be retained after the merger, and is a positive integer, measured in pieces. Indicates a small sample from the violation category A set of sample annotated evidence extracted from manually annotated legal provisions; Indicates based on the unit to be reviewed From the legal evidence database The collection of legal evidence retrieved; This function represents a merge function that removes duplicates by evidence number, sorts evidence by priority, and limits the number of evidence items.

[0053] Define the prosecution's set of doubts as Each point of doubt The valid indicator function for suspicious points is: (16); In equation (16), Indicates the unit to be reviewed A collection of points of contention raised by the prosecution; This represents the number of points of contention raised by the prosecution, and is a non-negative integer, expressed in units of [number]. Indicates the first A structured point of doubt, The range of values ​​is ; Points of doubt The validity indicator function takes the value 0 or 1; This represents the support score calculated according to formula (1); This indicates the support threshold, with a value range of [0,1] and no physical unit. The meaning is consistent with equation (15). The set of defense responses is defined as follows: , No. One defense response The function indicating successful rebuttal is: (17); In equation (17), Indicates the unit to be reviewed The set of defense responses; In response to doubts The response to the defense; Points of doubt Whether to respond to the defense The indicator function for a successful rebuttal, which takes the value 0 or 1; This represents the rebuttal score calculated according to formula (2); This represents the rebuttal threshold, with a value range of [0,1] and no physical unit. and The meaning is consistent with that of equation (16).

[0054] The asymmetric decision rule adopted in this invention is as follows: (18); In equation (18), Indicates the unit to be reviewed The asymmetric decision function takes the value 0 or 1; and The meaning is consistent with equations (16) and (17); as long as at least one valid point of contention is not successfully refuted, a non-compliant conclusion is output; the above decision is made by the rule engine based on the indicator function. and The value of is determined by calculation, and its result can be reproduced and verified under a given threshold and weight, without fluctuating with the single generation result of the large language model.

[0055] To describe the asymmetry of error cost, let... This indicates the cost of missed judgment. Indicates the cost of false alarms; in privacy compliance reviews ;like The output of non-compliance condition risk is: The risk of output compliance conditions is .in, Represents a conditional probability operator; This indicates the true compliance label of the unit under review. This indicates that the information is not compliant with regulations. This indicates authenticity and compliance; Indicates the unit to be reviewed. Represents a small sample case library This refers to a legal evidence repository. Therefore, Indicates that in a given unit to be reviewed Small sample case library and legal evidence database Under these conditions, the conditional probability of the unit under review being truly non-compliant, with a value range of [value missing], has no physical unit. When the risk of outputting non-compliance is no higher than the risk of outputting compliance, a non-compliance conclusion should be output, thus yielding the cost-sensitive optimal threshold: (19); In equation (19), This represents the cost-sensitive optimal threshold, with a value range of [0,1] and no physical unit. This indicates the cost of missed judgment. Both represent the cost of false alarms and are non-negative numbers under the same cost measurement system. Indicates the unit to be reviewed The conditional probability of true non-compliance ranges from [0,1]; because ,therefore The asymmetric adjudication mechanism of this invention engineeres the cost-sensitive decision-making concept into a structured evidence space. It should be noted that... It is a theoretically optimal decision threshold defined on the true probability of non-compliance, used to justify the cost-sensitive rationale of asymmetric adjudication. The conclusion in this invention is reached through the asymmetric decision function of equation (18). The structure (any valid, unrefuted objection is deemed non-compliant) is engineered, rather than defined through... or The specific numerical implementation. With definition in support of division Support threshold Definition in refutation The rebuttal threshold They belong to different dimensions and judgment stages, and their values ​​are not comparable.

[0056] To control for the potential increase in false alarms due to asymmetric rules, assume that the controlling party may raise a maximum of [number missing] false alarms. One point of doubt concerns the authenticity and compliance terms. The probability of any error or suspected point passing through the support gate does not exceed The probability that any gated error will be successfully refuted by the defense is no less than The probability that a single suspected error passes through the gating system simultaneously without being refuted does not exceed [a certain value]. The upper bound of the false alarm rate can be obtained from the union bound: (20); In equation (20), Represents the asymmetric decision function The false alarm rate, ranging from [0,1], has no physical units. This represents the upper limit of the number of points of suspicion raised by the prosecution; the value is a non-negative integer, and the unit is "points". This represents the upper bound of the probability that a suspected error in a genuine compliance clause will pass the support gate control, with a value range of [0,1] and no physical unit; This represents the lower bound of the probability that a gated error is successfully refuted by the defense. It takes a value in the range of [0,1] and has no physical unit. By merging legal provisions as evidence, filtering by legal provision number, retrieval based on chapter awareness, compliance protection rules, and lowering support thresholds; Through compliance-related small samples and topic perception scoring The constraints on the span of evidence and the citation capping mechanism have been improved; This can be reduced by limiting the number of points of contention raised by the prosecution and by merging duplicate points of contention.

[0057] (twenty one); In equation (21), Indicates at the candidate threshold The empirical asymmetric risk is given, and its value is a non-negative real number. This represents the verification buffer, where each element is a review record labeled with "human review". Indicates the number of records in the buffer, in units of records; This indicates a compliance label that has been manually reviewed; its value is either 0 or 1. This represents the system decision recalculated according to equation (18) under the candidate threshold, with a value of 0 or 1; This indicates the cost of missed judgment. This represents the cost of false alarms; both are non-negative numbers within the same cost measurement system. ; This indicates an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise.

[0058] (twenty two); In equation (22), This represents the self-correction cycle number, and its value is a non-negative integer. , Indicates the first Support thresholds and rebuttal thresholds used in the round; , Indicates the first The optimal candidate threshold is obtained on the round buffer according to equation (21); The learning rate is defined as a value within a certain range and has no physical unit. It is used to control the threshold update step size and self-correction speed. Equation (22) binds historical review records and manual review feedback to the threshold update process, so that the adjudication threshold is automatically adjusted according to the sample distribution and regulatory changes, rather than being a fixed constant.

[0059] Through the above mechanism, the adjudication threshold of this invention forms a closed-loop iteration with historical review records and manual review feedback: the system outputs the adjudication and writes it into the log, and the manual review result is written back and included in the next round buffer, triggering the threshold to be re-estimated under the asymmetric cost objective and self-corrected by exponential moving average, so as to automatically adjust the threshold when the sample distribution drifts, the violation type is updated or the regulations change, overcoming the problem that fixed thresholds are difficult to adapt to changes in data and regulations.

[0060] The technical solution of this invention will be described below in the context of a mobile application privacy policy compliance review scenario. This embodiment is for illustrative purposes only and does not limit the scope of protection of this invention. Figure 1 This is an overall architecture diagram of the privacy compliance judgment system in this embodiment of the invention, showing the data flow relationship between the modules of text preprocessing, declarative chapter filtering, investigation files, role-based small sample retrieval, legal evidence merging, prosecution doubts, support for gating, compliance protection, defense arguments, topic-aware rebuttal scoring, asymmetric adjudication, explanation of judgment reasons, and log auditing. Figure 2 This is a flowchart of the overall execution of the privacy compliance judgment method in this embodiment of the invention, showing the overall process from the input of the text to be reviewed to the output of the structured review log; Figure 3 This is a flowchart illustrating the process of generating investigation files in an embodiment of the present invention, showing the segmentation of the unit to be reviewed, the generation of basic metadata, and the creation of the investigation file. The process of extracting scene data, evidence fragments, inspection items, clues to doubts, and facts favorable to the defense; Figure 4 This is a flowchart illustrating the merging of role-based small samples and legal evidence in an embodiment of the present invention, showing small samples of violations. Compliance category small sample Legal Regulations Search Evidence Sample labeling evidence and the combined evidence set The generation process; Figure 5 This is a flowchart illustrating the generation of prosecution's doubts, filtering of legal statute numbers, and support for segmented gating in this embodiment of the invention. It shows the structured output of doubts, filtering of cited legal statutes, and support for segmented gating. Calculation and based on support threshold The effective process for determining doubts; Figure 6 This is a compliance protection flowchart in an embodiment of the present invention, illustrating the process of suppressing narrow-range false alarm patterns such as future recipients, statutory exceptions, and retention periods; Figure 7 This is a flowchart of the defense and rebuttal scoring process in an embodiment of the present invention, showing the structured output of the defense response. Six-tier scoring system, maximum evidence span, maximum total score without citations, and rebuttal threshold. Judgment process; Figure 8 This is an asymmetric formal adjudication flowchart in an embodiment of the present invention, showing the set of valid unrebutted objections. Construction and adjudication functions The rules for determining the value; Figure 9 This is a structured review log field relationship diagram in an embodiment of the present invention, showing the field relationships of input echo, investigation files, retrieval records, list of doubts, list of defenses, adjudication results, and running metadata written to the log. Figure 10 This is a comparison of the method of the present invention with random retrieval, similarity retrieval, using only legal evidence, symmetric debate, and ablation configuration on the main evaluation indicators. Figure 11 The recall rate was displayed. Rec False alarm rate FPR F1 score and total cost-sensitive risk With the rebuttal threshold A changing curve. Figure 12 This is a model network structure and data flow diagram in an embodiment of the present invention, showing the word segmentation embedding layer, multi-head self-attention encoding layer and pooling layer of the text encoding neural network that map text into semantic vectors and drive role-based few-sample retrieval, as well as the data flow relationship between the large language model decoding network and the deterministic rule engine used by the prosecution agent, the defense agent and the judge agent. Figure 13 This is a privacy compliance judgment state diagram topology diagram in an embodiment of the present invention, showing a directed state diagram and the execution flow between its nodes, consisting of a starting node, a route determination, an investigation file generation node, a declarative or administrative chapter skipping node, a role-based retrieval node, a prosecution doubt generation node, a defense rebuttal scoring node, an asymmetric adjudication node, a judgment reasoning explanation node, and a termination node.

[0061] Specifically, the model structure and data processing flow are as follows: The model structure and data processing flow of this invention are as follows: Figure 12As shown, the system comprises a text encoding neural network running on one or more processors of a computing device, a large language model decoding network used by the prosecution, defense, and adjudication agents, and a deterministic rule engine. The processors include a central processing unit and / or a graphics processing unit. The text encoding neural network and the large language model decoding network map natural language text into computable semantic vectors and structured intermediate data through forward propagation. The rule engine performs deterministic calculations on the intermediate data based on preset thresholds, weights, and indicator functions, and outputs the final ruling. These three components constitute a data processing pipeline from the input of the text to be reviewed to the output of the ruling.

[0062] The text encoding neural network is used to implement the text vector representation function defined in equation (11). It is a deep neural network with a Transformer encoder architecture, which, from the input side to the output side, includes a word segmentation and word embedding layer, a positional encoding layer, and so on. The network consists of stacked bidirectional multi-head self-attention encoding layers, feedforward network layers, and pooling layers; this network maps the word sequence of any text segment to a... In one embodiment, a dense semantic vector is generated and normalized. The value is 1024. The semantic similarity between the unit to be reviewed and each candidate sample is calculated using the cosine similarity of their semantic vectors according to formula (10). Based on this, the role-based small sample retrieval module recalls violation-type small samples and compliance-type small samples from the violation sample pool and the compliance sample pool, respectively. In one embodiment, the text encoding neural network adopts the BAAI / bge-m3 text vector model with XLM-RoBERTa as the encoder; in other embodiments, it may adopt a model fine-tuned with legal domain corpus that can map the input text to... Other text encoding models for dimensional normalized semantic vectors.

[0063] The prosecution agent, defense agent, and judge agent are each implemented by a large language model decoding network. This network is a Transformer neural network with a decoder architecture, which includes a word embedding layer, a stacked multi-head causal self-attention layer and a feedforward network layer, and an output projection layer. It generates structured data conforming to a preset JSON pattern in an autoregressive manner under structured constraint decoding. The above Transformer network structure, network forward computation, and autoregressive text generation method can all be implemented using existing large language model technologies in this field. The prosecution agent encodes the unit to be reviewed, a small sample of violations, the case file, and the set of legal evidence into an input word sequence. After forward propagation, it outputs a structured set of doubts containing the doubt number, violation type, doubt claim, legal basis number, and original evidence fragment. The defense agent, after inputting valid doubts, a small sample of compliance, and legal evidence, outputs a structured defense response containing the review topic, defense reasons, cited legal provisions and compliance sample numbers, and scores based on four dimensions: functional necessity, authorization and purpose clarity, user control, and minimum necessity and proportionality principle. The judge agent only generates natural language judgment reasons after the ruling is determined, without modifying the ruling. The three agents share the same basic large language model decoding network and are loaded with system prompt templates and output format constraints corresponding to the roles of prosecution, defense, and judge, respectively.

[0064] The rule engine is a deterministic program module running on a processor, used to calculate the intermediate input data according to preset formulas, thresholds, and logical rules. The rule engine itself does not perform random sampling based on probability distribution. The rule engine reads the semantic similarity output of the text encoding neural network and the structured doubts and structured defense responses output by the large language model decoding network. It calculates the support score of the doubts and the rebuttal score of the defense responses according to Equations (1) and (2), and calculates the validity of the doubts, the success of the rebuttal, and the asymmetric adjudication results according to the indicator functions of Equations (16), (17), and (18). Since the final adjudication is output by the rule engine after deterministic calculation of the above intermediate data, rather than directly given by the large language model decoding network, under the same threshold and weight, the same unit to be reviewed will obtain the same adjudication result after the same network forward propagation and rule engine calculation. This eliminates the conclusion fluctuations caused by the randomness of sampling and the sensitivity of prompt words in the large language model decoding network, and makes the adjudication results reproducible on the computing device and formally verified.

[0065] In one embodiment, the data processing flow of the present invention is arranged into a directed computation graph according to preset processing nodes and directed connection relationships between nodes, and deployed and run on a computing device, the topology of which is as follows: Figure 13As shown, the directed computation graph includes a starting node, an investigation file generation node, a role-based retrieval node, a prosecution's doubts generation node, a defense's rebuttal scoring node, an asymmetric adjudication node, a judgment reasoning explanation node, and a declarative or administrative chapter skipping node. After the investigation file generation node, a routing decision is set up to branch by chapter type. When the unit to be reviewed belongs to a declarative or administrative chapter, it proceeds to the declarative or administrative chapter skipping node and directly outputs a low-risk compliance conclusion; otherwise, it sequentially passes through the role-based retrieval node, the prosecution's doubts generation node, the defense's rebuttal scoring node, the asymmetric adjudication node, and the judgment reasoning explanation node before reaching the termination node. Intermediate data is passed between nodes via state objects, enabling the aforementioned text encoding neural network, large language model decoding network, and deterministic rule engine to be sequentially scheduled and executed by the processor according to a preset directed flow.

[0066] Example 1: Compliance assessment of the privacy policy paragraph; After the computing device obtains the full text of a mobile application's privacy policy, the text preprocessing module identifies the title, hierarchy, list number, and paragraph, and divides the full text into several units to be reviewed; each unit to be reviewed saves the policy identifier, paragraph number, chapter title, and paragraph text.

[0067] When the chapter title is a declarative or administrative chapter, the declarative chapter filtering module directly outputs a low-risk compliance result and records it in the log. This process avoids the controlling agent raising meaningless questions about paragraphs that do not describe specific data processing activities.

[0068] For substantive passages, the investigation file module first identifies the data processing scenario. For example, for a passage containing "We collect device IMEI, MAC address, and location information for risk control and anti-fraud," the investigation file module can identify scenarios involving sensitive personal information processing, device identifier processing, security risk control processing, and the consequences of authorization and denial; evidence fragments include "device IMEI," "MAC address," "location information," and "risk control and anti-fraud"; check items include whether the specific purpose is stated, whether the authorization method is stated, whether the consequences of denial are stated, whether the minimum necessity is stated, and whether the retention period is stated. The investigation file generation process is as follows: Figure 3 As shown.

[0069] The role-based small sample retrieval module then retrieves violation-related small samples based on paragraph text and chapter titles respectively. Compliance-related small samples If the paragraph to be reviewed is located in the "Collection and Use of Personal Information" section, the retrieval system will prioritize recalling samples from the same section or similar scenarios, excluding samples from the same policy. Samples of violations will only be entered into the prosecution's AI agent, while samples of compliance will only be entered into the defense's AI agent.

[0070] The legal evidence module performs a legal provision search based on the paragraph to be examined, and simultaneously reads... Manually annotated This serves as sample evidence. The two types of evidence are merged by number, deduplicated, and then input into the prosecution and defense. When the prosecution raises doubts, the legal basis number must exist in the merged legal evidence set; otherwise, that number is filtered out. The process for merging role-based small samples and legal evidence is as follows: Figure 4 As shown.

[0071] The prosecution's point of contention module generates structured points of contention based on paragraph text, case files, small samples of violations, and legal evidence. The module includes prosecution point of contention generation, legal provision numbering filtering, and support for gated processes. Figure 5 As shown. Taking the issue of "insufficient disclosure of sensitive personal information" as an example, if the legal provisions match... Similarity of non-compliant samples Text gap degree And weight , , ,but: ,in, , , , , , and The meaning is consistent with equation (1); This indicates that the legal provision matching degree in this embodiment is 0.9. This indicates a similarity of 0.8 between the non-compliant samples. This indicates a text gap level of 1.0; all values ​​have no physical units. (If thresholding is supported...) If a point of doubt is found to be true, it will be included in the set of valid points of doubt. If a point of doubt lacks legal support or original textual evidence, its support score will be below the support threshold, and it will not proceed to the next ruling.

[0072] The compliance protection module suppresses false alarms within a narrow scope, targeting future recipients, statutory exceptions, and retention periods. Its processing flow is as follows: Figure 6 As shown. For any doubts that violate compliance protection conditions, the system reduces their validity or requires them to obtain a higher support score before proceeding to the subsequent rebuttal and adjudication stages.

[0073] The defense's rebuttal module receives valid points of contention, paragraph texts, compliance-related small samples, and legal evidence, and generates a rebuttal response point by point. For each point of contention, the defense... , , , The scoring is based on four dimensions, with each evaluation item preferably taking a value from a set of six anchor points {0, 0.2, 0.4, 0.6, 0.8, 1}. If a dimension does not provide a corresponding original text evidence span, the score for that dimension can only reach the first preset upper limit; if the entire defense neither cites legal provisions nor compliant samples, the total rebuttal score can only reach the second preset upper limit. The scoring process for the defense's arguments and rebuttals is as follows: Figure 7 As shown.

[0074] The rebuttal scoring module selects weights based on the topic being reviewed. For example, in the topic of user rights, The weight of this dimension is higher than other dimensions; in the topics of storage, third-party sharing, and cross-border transmission, The weighting is relatively high; in the topic of security measures, and The dimension weight is relatively high. If a rebuttal is made... Greater than If the statement is true, then the point of contention has been successfully refuted; otherwise, it is considered unrefuted.

[0075] The asymmetric adjudication module reads valid objections and rebuttals. If all valid objections are successfully refuted, compliance is output; if any valid objection is not successfully refuted, non-compliance is output, and that objection is added to the unresolved objection set. The adjudication reasoning module then generates a natural language explanation based on the established conclusion, which must not change the adjudication result. The asymmetric formal adjudication process is as follows: Figure 8 As shown, the field relationships for input echoes, search records, lists of doubts, lists of defenses, rulings, and runtime metadata written to the structured review log during the review process are as follows: Figure 9 As shown.

[0076] Example 2: Review of Third-Party Sharing Terms; In another embodiment, the paragraph to be examined describes "we may share necessary information with our partners in order to provide related services." The investigation file module identifies that this paragraph involves a third-party sharing scenario and extracts "partners," "necessary information," and "related services" as evidence fragments. Because the above statement may have issues such as ambiguity regarding the recipient, processing purpose, processing method, and type of personal information, the prosecution's suspicion module can raise suspicion of "insufficient disclosure of third-party sharing."

[0077] If the legal evidence collection contains clauses stipulating that when providing personal information to other personal information processors, the recipient's name or contact information, the purpose of processing, the method of processing, the type of personal information, and consent must be obtained, then the doubtful point... High. If Similar violations exist in China. Higher. If the original text indeed does not specify the recipient and data type, higher. If so, the suspicious point will be detected through the gate control.

[0078] The defense module then used this paragraph and a small sample of compliance cases as a basis for its arguments. Generate a structured defense response. If the paragraph explicitly lists the recipient, type of personal information, purpose of use, processing method, user control path, or authorization method, the defense may... , , High scores are given across three dimensions, with excerpts from the original text serving as the evidence span; if the paragraph does not adequately disclose the above elements, the rebuttal score will be below the threshold. If the doubts are not successfully refuted, the output of the asymmetric adjudication module is non-compliant.

[0079] Example 3: Review of User Rights Terms; In the section on user rights, the paragraph awaiting review may describe, "You can contact us through customer service to handle requests related to your personal information." The prosecution may raise concerns about the "unclear mechanism for realizing user rights." This is because user control is a key issue under this topic. Compliance defenses are more important; the rebuttal scoring module selects weights corresponding to user rights topics, making... The weight of this dimension is higher than that of other dimensions.

[0080] If the original text provides specific paths for querying, copying, correcting, deleting, withdrawing consent, and canceling the account, and specifies the response time or contact information, then... The score is relatively high. or We can assign a score based on the adequacy of the explanation. If the original text only vaguely states "Contact customer service" without providing a path, method, or timeframe, then... The score is low, and the doubts are difficult to refute successfully.

[0081] To verify the effectiveness of this invention, experiments were conducted on a Chinese mobile application privacy policy compliance review dataset. The experimental data included multiple mobile application privacy policies, which were segmented by paragraph, clause, and semantic boundaries to obtain review units. Each review unit was manually labeled with compliance tags, violation types, and applicable legal provisions. The small sample case library consisted of a violation sample pool and a compliance sample pool. The legal evidence library covered laws and regulations related to personal information protection, normative documents, and a manually maintained violation pattern library. The aforementioned small sample case library and legal evidence library were obtained through the following methods: The original texts of publicly released Chinese mobile application privacy policies, user agreements, and other personal information processing documents were collected and segmented into review units according to paragraphs, clauses, headings, or semantic boundaries; each review unit was manually annotated with compliance tags, violation types, applicable legal provision numbers, expert reasoning text, and chapter titles or review scenario tags by legal professionals, and then entered into the library after cross-verification for consistency checks; samples were further divided into a retrieval pool and a test set according to their respective policies, with samples from the same policy not crossing sets, to avoid data leakage caused by recalling similar neighbors from the same policy during the retrieval stage. Samples marked as non-compliant constituted the violation sample pool, and samples marked as compliant constituted the compliance sample pool. The legal evidence database consists of provisions from laws and regulations such as the Personal Information Protection Law, the Cybersecurity Law, the Data Security Law, and methods for identifying illegal and irregular collection and use of personal information by apps, as well as national standards, departmental rules, regulatory notices, administrative penalty cases, and judicial rulings. These provisions are divided into blocks according to the provision number, law name, clause number, and category. A mapping database between violation patterns and recommended provision numbers is manually maintained by domain experts, allowing for both violation pattern-based indexing and vector retrieval.

[0082] In one experiment, the method of this invention was compared with Zero-Shot (zero-shot learning), random functions, similarity sim, RAG (Retrieval-Augmented Generation) only, symmetric debate, and several ablation configurations. The symmetric debate scheme adds a judge's large language model to the statements of both the prosecution and the defense to comprehensively determine compliance, and tends to output a compliance conclusion when evidence is insufficient or uncertain. Experimental results show that on 247 test paragraphs, the method of this invention has an accuracy of 0.798, precision of 0.716, recall of 0.910, F1 score of 0.802, and false positive rate of 0.294; the symmetric debate scheme has an accuracy of 0.688, precision of 0.840, recall of only 0.378, F1 score of 0.522, and false positive rate of 0.059; the Zero-Shot scheme has an F1 score of 0.422; and the RAG only scheme has an F1 score of 0.613. The results indicate that role-based evidence retrieval, legal evidence, defense rebuttal, and asymmetric adjudication collectively improve overall judgment. Figure 10The results show a comparison of the proposed method with random retrieval, similarity retrieval, legal evidence-only, symmetric debate, and ablation configuration on key evaluation metrics. Compared to the control methods, the proposed method achieves the highest levels in precision, recall, and F1 score. The symmetric debate scheme, due to the introduction of a judge's large language model that tends to rule in compliance when evidence is insufficient, has a low false positive rate (0.059), but its recall rate drops significantly to 0.378, indicating serious missed judgments, and its F1 score is only 0.522. The proposed method, through asymmetric adjudication and evidence constraints, achieves the highest recall rate (0.910) and comprehensive F1 score (0.802) while maintaining a low false positive rate. In contrast, the symmetric debate scheme, due to the judge's large language model's tendency to rule in compliance when uncertain, has significantly lower recall and F1 scores than the proposed method. This demonstrates that the proposed method is significantly superior to the symmetric adjudication-based aggregation in its violation identification capability and achieves better results in suppressing missed judgments.

[0083] Table 1. Main experimental results of the method of the present invention and some comparative methods. ; In the ablation experiments, after removing the small number of compliance samples, the F1 score decreased from 0.802 to 0.751; after removing legal evidence, the system could not form valid doubts, and the F1 score dropped to 0; after canceling role-based retrieval, the F1 score dropped to 0.727. These results indicate that compliance samples, legal evidence, and role-based segregation retrieval all contribute to system performance.

[0084] Table 2 Comparison of results of the method of the present invention with several ablation configurations. ; When legal evidence is removed, the system outputs a compliance conclusion for all units under review. In this case, the accuracy of 0.551 equals the proportion of truly compliant segments in the test set, i.e., the accuracy of the trivial baseline of "all judged as compliant," and does not reflect any violation identification capability. Under this configuration, precision, recall, and F1 are all 0, indicating that the system has lost its discriminative ability. Therefore, accuracy alone cannot reflect the effectiveness of compliance review on datasets where compliant samples constitute the majority, while recall, F1, and false positive rate better reflect the true discriminative ability. This is why this invention uses F1 and false positive rate as the main evaluation indicators.

[0085] In another experiment, for and An estimate is made. Under the method of this invention, the probability of an erroneous point passing through the gating system is... The estimated value is 0.537 (based on the proportion of clauses with at least one valid error in a genuine compliance clause), representing the probability that a gated error will be successfully refuted by the defense. The estimated value is 0.433 (when the rebuttal threshold is 0.75). This result indicates the probability that a valid erroneous point of contention in a single clause will not be refuted. The false alarm rate is approximately 0.304, which is close to the measured false alarm rate of 0.294 for the method of this invention, indicating that the support for gating and the defense rebuttal mechanism can jointly constrain the upper limit of false alarms. .

[0086] (twenty three); In equation (23), Indicates the threshold for rebuttal. The total cost-sensitive risk at that time is a non-negative real number with no physical unit. Indicates the threshold for rebuttal. The recall rate of the system at that time, with a range of values; Indicates the threshold for rebuttal. The false alarm rate of the system is within a certain range. Indicates the false negative rate; This indicates the cost of missed judgment. This represents the cost of false alarms; both are non-negative numbers within the same cost measurement system, and... Equation (23) obtains the total cost-sensitive risk that varies with the rebuttal threshold by weighting the false positive rate and the false negative rate according to asymmetric costs. This makes the optimal value of the rebuttal threshold fall at the minimum point of the total risk, reflecting the asymmetric cost of false positives being higher than that of false negatives in privacy compliance review.

[0087] Figure 11 It displays the recall rate (Rec), false positive rate (FPR), F1 score, and total cost-sensitive risk. With the rebuttal threshold The curves represent the changing parameters, with the left vertical axis representing recall (Rec), false positive rate (FPR), and F1 score, and the right vertical axis representing total cost-sensitive risk. .

[0088] The total cost-sensitive risk is calculated according to formula (23), which is the ratio of the cost of missed detection to the cost of false alarm in a preferred embodiment. The ratio is set to 1.5:1. While maintaining the support for segmented gating, the ratio increases with the rebuttal threshold. As the number of cases increases, the probability of false alarms being successfully refuted by the defense decreases, the false alarm rate Rec increases, and the recall rate Rec improves; after asymmetric cost weighting by equation (23), the total cost-sensitive risk... It shows a trend of first decreasing and then increasing. For example... Figure 11 As shown, exist The minimum value is obtained at the point where the cost-sensitive optimal threshold idea described in equation (19) is reflected in the rebuttal threshold. Near this preferred threshold, the system can retain the recall rate Rec while suppressing erroneous doubts from entering the final non-compliant conclusion. Therefore, the preferred rebuttal threshold of this invention is... .

[0089] In a preferred embodiment, a threshold is supported. Set to 0.5, the rebuttal threshold. Setting it to 0.75 supports weighted distribution. , , The values ​​were set to 0.5, 0.3, and 0.2 respectively for small samples of the violation category. Set to 3, for a small sample size of compliance-related samples. Set to 3, for legal evidence Set to 5. Because it supports points. After filtering by legal provision number, matching compliance protection rules and legal provisions, similarity of violation samples, and weighting by gaps in the original text, the support score distribution of valid doubts is generally high, with a median of approximately 0.9 in one embodiment. Therefore, the support threshold is... Within a relatively wide value range, the impact on the final decision is small, and system performance has little effect on the support threshold. Insensitive; while rebuttal threshold The success of rebuttal significantly impacts whether doubts are refuted and the final ruling; therefore, this invention addresses the rebuttal threshold. The analysis focuses on the changes in the value of the threshold, refuting the previous conclusions. The effect curve of the change on the system's false alarm control capability is as follows: Figure 11 As shown.

[0090] In a preferred embodiment The scoring set is {0, 0.2, 0.4, 0.6, 0.8, 1}. When a dimension lacks original evidence, the score for that dimension is capped at 0.4; when the defense response does not cite legal evidence or compliant samples, the rebuttal score is capped at 0.6.

[0091] In a preferred embodiment, each review topic Corresponding weight vector The settings are shown in the table below, where , , , Functional necessity Authorization and Clarity of Purpose User control rights and the principle of minimum necessity and proportionality The weights of each theme are such that the sum of the four-dimensional weights of each theme is 1.

[0092] Table 3 shows the weight parameters corresponding to different themes in the preferred embodiments. ; Table 4. Main parameter names and values ​​in the preferred embodiments ; The aforementioned threshold is an initial preferred value, determined by grid search on historical labeled data. During the continuous operation of the system, the initial threshold is periodically re-evaluated and self-corrected by exponential moving average based on the manual review feedback in the structured review log, according to the mechanism described in step 17.

[0093] The vector representation model in this invention can be a general Chinese vector model, a Chinese legal vector model, a domain-specific fine-tuned vector model, or other models capable of calculating text similarity. The legal evidence database can include laws and regulations, national standards, departmental rules, regulatory notices, administrative penalty cases, judicial judgments, internal corporate compliance systems, or manually maintained violation pattern libraries. The large language model interface can be a locally deployed model, a cloud-based model, an open-source model, or a private model. The aforementioned vector representation model is a text encoding neural network, which can be a general Chinese vector network with a Transformer encoder architecture, a Chinese legal domain-specific fine-tuned vector network, or other deep neural networks capable of calculating text similarity. The aforementioned prosecution agent, defense agent, and judge agent are implemented by a large language model decoding network. Different role system prompts and decoding parameters can be loaded onto the same basic decoding network, or different weight networks obtained through instruction fine-tuning, or neural networks of different sizes and architectures. For example, an encoder architecture network can be used in the retrieval stage, while a decoder architecture network can be used in the generation stage. Both the text encoding neural network and the large language model decoding network are deployed on the central processing unit or graphics processing unit of the computing device to perform forward propagation. The formulas described in this invention are deterministically calculated by the rule engine on the processor from the intermediate data output by the network, rather than being purely mathematical methods detached from the aforementioned networks and hardware.

[0094] This invention is not limited to privacy policy paragraphs, but can also be applied to rules for children's privacy protection, data sharing agreements, instructions for data export, rules for personal information processing, instructions for data classification and grading, instructions for de-identification rules, instructions for financial business data processing, and instructions for medical data processing. The method described in this invention can be used as long as the task has a structure involving evidence retrieval, raising of doubts, evaluation of defenses, and adjudication of asymmetric risks.

[0095] The present invention also provides a computer-readable storage medium storing computer instructions, which, when executed, cause the computer to perform the steps of the above-described privacy compliance judgment method based on role-based asymmetric adjudication.

Claims

1. A privacy compliance judgment method based on role-based asymmetric adjudication, wherein the method is executed by a computing device, the computing device including a processor, a text input interface to be reviewed, a legal evidence library, a small sample case library, a vector retrieval module, and a large language model interface, the method comprising: Acquisition steps: Acquire the text to be reviewed through the text input interface, and preprocess the text to be reviewed to extract at least one unit to be reviewed; Search steps: Based on the unit to be reviewed, semantic similarity search is performed in the legal evidence database and the small sample case database through the vector retrieval module to obtain the legal provision-related evidence set and case-related evidence set corresponding to the unit to be reviewed; Integration Step: Input the unit to be reviewed, the legal provision-related evidence set, and the case-related evidence set into the large language model interface, and generate a ruling result that integrates the legal provision-related evidence set and the case-related evidence set through the large language model; The method is characterized in that it further includes: The text to be reviewed is segmented into units and its metadata is annotated to obtain at least one unit to be reviewed; For each of the aforementioned entities to be examined, a role-based adversarial process and asymmetric adjudication process are executed, the process including: The sub-step for generating investigation files is as follows: Generate case file data for the current unit to be reviewed, including the processing scenario, fragments of original evidence, a set of inspection results, clues to doubts, facts favorable to the defense, and questions to be asked by the prosecution. Role-based retrieval sub-step: Retrieve violation-type small samples and compliance-type small samples from the small sample case library respectively; Evidence aggregation sub-step: Retrieve legal evidence from the legal evidence database and merge it with the sample-labeled evidence extracted from the small sample of violations to form a legal evidence set; The prosecution's sub-step involves inputting the current unit to be reviewed, the small sample of violations, the legal evidence set, and the case file data into the prosecution's intelligent agent to generate at least one structured point of doubt. The prosecution's intelligent agent is implemented by a large language model decoding network, and the legal basis it cites must belong to the legal evidence set. The sub-step for determining the validity of doubts is as follows: a support score is calculated for each structured doubt. The support score is used to characterize the comprehensive support of the doubt in three dimensions: legal basis, violation cases, and text gaps. The valid doubts are determined based on the comparison result between the support score and the preset support threshold. The defense's rebuttal sub-step involves inputting the valid doubts, compliance-related small samples, legal evidence sets, and case file data into the defense agent to generate a rebuttal response for each valid doubt. The rebuttal response includes at least the doubt number, the subject of review, the defense arguments, the original text, the cited legal provision number, the cited compliance sample number, the uncovered prosecution points, and four quantitative evaluation items. These four quantitative evaluation items are: functional necessity evaluation item, authorization and purpose clarity evaluation item, user control evaluation item, and minimum necessity and proportionality principle evaluation item, respectively used to characterize the degree of compliance of the unit under review in terms of functional necessity, authorization clarity, user control, and proportionality principle. The sub-step for determining the validity of a rebuttal is as follows: Based on the review topic to which the current unit to be reviewed belongs, assign corresponding topic weights to the four quantitative evaluation items, calculate the rebuttal score by weighting, and determine whether the doubt has been successfully refuted based on the comparison result between the rebuttal score and the preset rebuttal threshold. Asymmetric adjudication sub-step: By using the rule engine running on the processor, valid doubts that were not successfully refuted are counted. If there is at least one valid doubt that was not successfully refuted, a non-compliant conclusion is output; otherwise, a compliant conclusion is output. The adjudication result is not directly generated by the large language model. The output step involves inputting the ruling result, valid doubts, defense responses, unresolved doubts, and rebuttal points into the adjudicator's intelligent agent to generate the adjudication reasoning in natural language.

2. The privacy compliance judgment method based on role-based asymmetric adjudication according to claim 1, characterized in that, The sub-step for generating the investigation file also includes: Based on the chapter title or paragraph topic of the unit to be reviewed, determine whether it belongs to a declarative or administrative chapter; If so, skip the subsequent prosecution and defense process, directly output a low-risk compliance conclusion, and record the reason for skipping in the review log; If not, then the case file data is generated.

3. The privacy compliance judgment method based on role-based asymmetric adjudication according to claim 1, characterized in that, In the role-based retrieval sub-step: The violation-related small samples are retrieved only from the sample pool marked as non-compliant in the small sample case library; The compliance-related small samples are retrieved only from the sample pool marked as compliant in the small sample case library; Furthermore, when performing the comparison document screening, samples with the same chapter title or the same review scenario as the current unit under review are selected as the first comparison benchmark; interference samples with the same policy source identifier as the current unit under review are excluded from the first comparison benchmark, and the target comparison document set is constructed from the remaining samples other than the interference samples.

4. The privacy compliance judgment method based on role-based asymmetric adjudication according to claim 1, characterized in that, The support score is calculated in the following way: For each structured point of doubt, calculate its legal provision matching degree with the legal evidence set, its case similarity with the small sample of violations, and its text missing intensity in the current unit to be reviewed; The support score is obtained by weighted summation of the legal provision matching degree, case similarity, and text missing intensity; the weights of the legal provision matching degree, case similarity, and text missing intensity are all preset non-negative values, and the sum of the three is 1; The calculation method for the legal provision matching degree is as follows: Obtain the legal basis number cited for the structured doubts; Search for the legal citation scores corresponding to each legal basis number in the legal evidence set, take the highest score and normalize it, and use the normalized value as the legal citation matching degree; If the legal basis number cited in the structured doubt is not in the set of legal evidence, then the legal provision matching degree is 0.

5. The privacy compliance judgment method based on role-based asymmetric adjudication according to claim 4, characterized in that, The rebuttal score is calculated as follows: Based on the review topic to which the current unit to be reviewed belongs, determine the four topic weights that correspond one-to-one with the four quantitative evaluation items; The rebuttal score is obtained by multiplying the value of each quantitative evaluation item by its corresponding topic weight and then summing the results. The values ​​of the four quantitative evaluation items are all selected from a set of six preset anchor points. The input prompts of the defense agent contain preset scoring specifications for each evaluation dimension and six anchor point judgment criteria. The defense agent outputs the values ​​of each quantitative evaluation item, the span of the original text evidence on which it is based, and the reasons for the score according to the specifications, and writes them into the structured review log.

6. The privacy compliance judgment method based on role-based asymmetric adjudication according to claim 1, characterized in that, In the defense sub-step, evidentiary constraints are applied to the defense response: If no corresponding original evidence span is given for any quantitative evaluation dimension, the score for that dimension will be capped at the first preset upper limit. If the defense response does not cite any legal evidence or any compliance-related small samples, then the rebuttal score will be capped at the second preset upper limit; If the defense response lists prosecution points that are not covered, the rebuttal score will be deducted according to a preset deduction coefficient.

7. The privacy compliance judgment method based on role-based asymmetric adjudication according to claim 1, characterized in that, In the asymmetric decision sub-step: The rule engine reads the semantic similarity data output by the vector retrieval module, the structured doubt data output by the prosecution agent, and the structured defense response data output by the defense agent. The rule engine outputs the final ruling result according to the structured question data and the structured defense response data, using deterministic calculation logic, so that the same input data, calculated by the rule engine under the same parameters, yields the same ruling result.

8. The privacy compliance judgment method based on role-based asymmetric adjudication according to claim 1, characterized in that, The method further includes a dynamic self-correction step for the threshold: The computing device maintains a verification buffer for storing historical review records with manual review labels; When the number of labeled records in the verification buffer reaches a preset lower limit, multiple sets of candidate support thresholds and candidate rebuttal thresholds are traversed in the candidate threshold grid, and the empirical asymmetric risk under each set of candidate thresholds is calculated respectively. Select a set of candidate thresholds that minimizes the empirical asymmetric risk as the optimal threshold for this round; The current threshold is updated using an exponential moving average with a preset learning rate, and the updated threshold is then clipped to a preset range.

9. The privacy compliance judgment method based on role-based asymmetric adjudication according to any one of claims 1 to 8, characterized in that, The prosecution, defense, and referee agents are each implemented by a large language model decoding network. The large language model decoding network is a Transformer neural network with a decoder architecture, which generates candidate outputs in an autoregressive manner. The fields and formats of the candidate outputs are limited by preset output rules in the input prompts, and the processor performs structural parsing and field verification to form structured data that conforms to the preset format. The vector retrieval module is implemented by a text encoding neural network with a Transformer encoder architecture, which is used to map text fragments into dense semantic vectors.

10. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed by the processor, the processor performs the steps of the privacy compliance judgment method based on role-based asymmetric adjudication as described in any one of claims 1 to 8.

Images