Artificial intelligence dynamic audit process arrangement method and system based on risk score

Abstract

The application provides an artificial intelligence dynamic auditing process arrangement method and system based on risk score, and belongs to the technical field of intelligent content auditing. The method is executed by a computer system: multi-modal analysis is performed on an object to be audited to obtain a dimension risk score; a comprehensive uncertainty is calculated from a category distribution output by an auditing model, and a comprehensive risk score obtained by weighted summation is combined to synthesize an auditing intensity requirement; an interception rate parameter and a resource occupation parameter are configured for a candidate auditing node, and a process intensity is aggregated and calculated; a target auditing process is solved with the process intensity satisfying an auditing intensity requirement and a preset hard rule as constraints, and the total amount of auditing resource occupation being minimized as a target; the auditing intensity requirement is continuously monitored and recalculated during execution, and a remaining sub-process is automatically recalculated according to a remaining intensity requirement; and parameters are calibrated according to an artificial auditing conclusion. The application dynamically matches the auditing process and the content risk state, and reduces the processing time delay and the auditing resource occupation.

Description

Technical Field

[0001] This invention relates to the field of intelligent content review technology, and in particular to an artificial intelligence dynamic review process arrangement method and system based on risk scoring. Background Technology

[0002] With the continuous growth in the scale of news media, government announcements, and content platform releases, the review of texts, images, videos, and other content has become an indispensable part of the release process. Currently, review systems generally adopt an architecture of "detection model + manual review process": first, text review models, image review models, etc., detect the content, and then reviewers complete the review according to a pre-configured process. The existing review process organization methods mainly fall into the following categories: (1) Fixed template process. The review process is pre-configured by the operations staff as a fixed template. Regardless of the risk characteristics of the content to be reviewed, it flows along the same path. Some systems have set up simple branches for machine scoring on the basis of the fixed process. For example, content with a score exceeding the preset threshold is transferred to manual review.

[0003] (2) Multimodal detection results are presented side by side. The detection of text, images and videos is completed by independent models. The detection results are presented to the reviewers in a list format, and the reviewers can make their own judgment on the subsequent processing method based on the results of each modality.

[0004] (3) Summary of conclusions from multiple models. Multiple detection models are deployed for the same content, and a comprehensive conclusion is output by means of voting or weighting. The model confidence is only used for interface display.

[0005] The problems that exist in the operation of the above solution include: 1. The review process is statically tied to content risk, resulting in unreasonable overall system processing delays and resource consumption. Regardless of the content risk level, all content goes through the same review stage, causing a large number of low-risk contents to queue up, consuming review resources and prolonging the release cycle; while a small number of high-risk contents cannot obtain in-depth review commensurate with their risk level, resulting in a long-term mismatch between processing delays and risk status.

[0006] 2. There is a lack of a calculable unified metric among the multi-source detection results. The detection results of text, images, and videos are heterogeneous in form and scale. The system can only present them side by side, but cannot convert them into quantities that can be used for process calculations. Process decisions still rely on manual completion and cannot achieve automated processing.

[0007] 3. Fixed threshold branches cannot adapt to runtime changes. The process branch thresholds are set manually and remain unchanged for a long time. When the content is modified during process execution, the scope of propagation expands, or the external environment changes, the generated process cannot be adjusted accordingly and can only rely on manual reconfiguration. The response is significantly delayed, and the adjustment process is difficult to trace.

[0008] In addition, existing technologies also have problems such as insufficient confidence of the audit model or conflicting conclusions of different models, lack of corresponding process handling mechanisms, model misjudgment cannot be compensated through process steps, and manual review conclusions, rejection reasons and release results are not used to correct detection parameters and process thresholds. After the distribution of business data changes, the detection results and process configuration continue to deviate from the actual risk state, and the performance gradually degrades.

[0009] In summary, due to the overlapping of the above problems, the existing review system has long been in a state of mismatch between review processing delay, review resource allocation and the actual risk status of the content, and is unable to adapt to changes in risk status during process execution, making it difficult for the system to simultaneously meet the requirements of release timeliness and review quality. Summary of the Invention

[0010] To address the aforementioned technical issues, this invention provides an AI-based dynamic review process orchestration method and system based on risk scoring. It aims to resolve problems in existing review systems such as static binding of review processes to content risk status, inability to uniformly quantify multi-source detection results, inability of fixed-threshold processes to adapt to runtime changes, lack of utilization of model uncertainty information, and lack of review result feedback. This allows the generation and adjustment of the review process to match the actual risk status and dynamic changes of the content to be reviewed, reducing overall system processing latency while meeting review quality requirements, enabling on-demand allocation of review resources, and maintaining long-term system effectiveness under changing business environments through continuous feedback of review results.

[0011] To achieve the above objectives, this application proposes an AI-based dynamic review process orchestration method based on risk scoring, executed by a computer system. The computer system is communicatively connected to multiple review models and multiple review terminals, and includes: S1: Receive the object to be reviewed and its business context, perform multimodal analysis on the text, image, video and audio data of the object to be reviewed, and obtain dimensional risk scores for multiple risk dimensions; S2: Perform probability calculation on the category distribution output by the audit model to obtain the comprehensive uncertainty; sum the dimensional risk scores according to the scenario weights to obtain the comprehensive risk score; calculate the audit intensity requirement based on the comprehensive risk score and the comprehensive uncertainty adjusted by the uncertainty penalty coefficient. S3: Configure interception rate parameters and resource consumption parameters for candidate review nodes, and aggregate and calculate the process strength based on the interception rate parameters of each review node in the candidate review process. The process strength increases monotonically with the increase of the number of review nodes participating in the aggregation and does not exceed 1. S4: Based on the constraint optimization algorithm, among the candidate review processes whose process intensity is not less than the review intensity requirement and meet the preset hard rules, the target review process is automatically solved with the goal of minimizing the total amount of review resources occupied, and the review tasks are automatically distributed to the review terminals corresponding to each review node according to the target review process. S5: During the execution of the target audit process, continuously acquire monitoring data associated with the object to be audited and recalculate the audit intensity requirement. When the preset adjustment conditions are met, determine the remaining intensity requirement based on the process intensity of the executed audit nodes, and automatically re-solve the remaining sub-processes that have not yet been executed in order to adjust the target audit process. S6: Collect the conclusions of manual review, and based on the deviation between the conclusions of manual review and the corresponding review intensity requirements, update the scene weight, intensity threshold and interception rate parameters according to the preset calibration loss function, and feed them back to S2 to S5.

[0012] As a further solution, the comprehensive uncertainty is obtained by weighted combination of model output uncertainty, inter-model divergence, and missing dimension penalty term; wherein, the model output uncertainty is obtained by weighted summation of the normalized entropy of the category distribution output by each review model according to modal weights, the inter-model divergence is the mean of the difference between the category distribution output by any two review models, and the missing dimension penalty term is the sum of the scenario weights corresponding to the missing risk dimension.

[0013] As a further solution, solving the target review process also includes: The review intensity level is mapped according to the intensity range into which the review intensity requirement falls; when a key risk label is hit, the review intensity level is constrained to be no lower than the minimum review intensity level corresponding to that label; When the overall uncertainty exceeds the uncertainty threshold, or the difference between the category distributions output by any two audit models exceeds the divergence threshold, the manual review node is added to the target audit process as a mandatory audit node, and the conclusions, confidence levels, and hit text fragments, video frames, or source annotations of each audit model are sent to the audit terminal corresponding to the manual review node along with the audit task.

[0014] As a further solution, the interception rate parameter is the probability that the review node independently intercepts illegal content. Its initial value is preset according to professional division of labor and historical review performance, and after running, it is automatically updated according to the ratio of the number of illegal content actually intercepted and confirmed in the sliding window to the total number of illegal content entering the review node and confirmed. When there is a joint review group with multiple reviewers conducting parallel reviews, the joint review group participates in the process strength calculation as an equivalent review node, and its equivalent interception rate is higher than the interception rate parameter of any reviewer in the joint review group; the preset hard rules include review nodes and tenant business rules that must be included or prohibited from being included; the calculation of the total amount of review resources also includes a latency item.

[0015] As a further solution, the preset adjustment conditions include: recalculated review intensity requirements. The difference between the audit intensity requirement at the time of arrangement and the incremental threshold is exceeded, the audit intensity level is crossed, the newly added key risk label is hit, or the audit object is resubmitted with a new version; When the preset adjustment conditions are met, the remaining strength requirements that the remaining sub-processes need to meet decrease as the process strength of the executed audit nodes increases, and the remaining sub-processes are automatically re-solved accordingly. when If the difference between the audit intensity requirement during the arrangement exceeds the return threshold, the object to be audited will be returned to the previous audit node for re-audit with attached risk evidence. when When a decrease leads to a downgrade in the review intensity level, cancel any review nodes that have not yet been started and whose resource usage parameters are higher than the preset resource threshold. Each time the target audit process is adjusted, the preset adjustment conditions, risk increments, and the audit nodes that are added or canceled, along with their risk basis, are written into the audit chain record. This is used to trace back the source of the adjustment to the target audit process by audit node.

[0016] As a further solution, the audit intensity requirement is decomposed into dimensional audit intensity requirement components according to the risk dimension. Each requirement component increases with the increase of the dimension risk score and dimension uncertainty, and the influence of dimension uncertainty decreases with the increase of the dimension risk score. The sum of each requirement component is the audit intensity requirement. The dimensional uncertainty consists of the entropy contribution of the category group mapped to the risk dimension in the category distribution output by each audit model, the component mapped to the risk dimension in the inter-model divergence, and the missing dimension penalty term of the risk dimension; a dimensional interception capability vector is configured for the candidate audit node; Using the capability subspace as the constraint unit, the sum of each requirement component mapped to the same capability subspace is taken as the subspace strength requirement. The subspace process strength on each capability subspace is calculated by aggregating the interception capabilities of each review node. The target review process is solved with the constraint that the subspace process strength of each active subspace is not less than the corresponding subspace strength requirement. The capability subspace is formed by aggregating multiple risk dimensions according to professional division of labor, and the active subspace is the subspace whose intensity requirement exceeds the dimension participation threshold.

[0017] As a further solution, when solving the target review process, the strength gap of each active subspace is calculated, and candidate review nodes with higher dimensional interception capabilities and lower resource consumption parameters on unclosed strength gaps are preferentially selected to join the target review process until all strength gaps are closed. When a countersigning group exists, select a combination of complementary capabilities from each active subspace; if there is still an unclosed strength gap after traversing the available candidate review nodes, automatically insert an external submission node. During execution, if the current review node's dimensional interception capability in a certain active subspace is lower than the lower limit of its capability, and there is an idle candidate review node whose dimensional interception capability in that subspace is higher than the current review node's preset capability increment, then the review will be automatically transferred to that idle candidate review node. After execution, the dimensional interception capability of each review node is automatically updated subspace by subspace according to the ratio of the number of content that is actually intercepted and confirmed to be in violation within the sliding window to the total number of content that enters the review node and is confirmed to be in violation in the corresponding capability subspace. When a capability subspace fails to meet the constraints but still misses a process if it exceeds a preset number of times within a preset time window, the dimensional interception capability of each audit node in the subspace is reduced by a preset shrinkage factor.

[0018] As a further solution, the determination of the preset adjustment conditions is based on time extrapolation requirements, and the calculation of the time extrapolation requirements includes: Maintain the risk sampling sequence and estimate the risk drift rate based on the changes in the review intensity requirements at adjacent sampling time points; update the system-level benchmark deviation based on the time-weighted average of the difference between the quantified value of the manual review conclusion and the corresponding review intensity requirements, and only include the time extrapolation requirement when the system-level benchmark deviation exceeds the preset control band; The time extrapolation requirement is the sum of the audit intensity requirement at the current sampling time, the product of the risk drift rate and the remaining processing time after truncation, and the system-level benchmark deviation. The audit intensity level determination and the judgment of the preset adjustment conditions are performed by replacing the recalculated audit intensity requirement with the time extrapolation requirement. When the time extrapolation requirement exceeds the intensity threshold of the current review intensity level, the review process is upgraded in advance; when the remaining processing time is less than the delay required for adding a new review node, the release of the object to be reviewed is suspended until the supplementary review is completed. When the time extrapolation requirement is lower than the current review intensity level lower limit and continues for a preset duration, cancel the review node whose resource usage parameter is higher than the preset resource threshold that has not yet been started; shorten the risk sampling cycle for the pending review object whose absolute value of risk drift rate exceeds the drift concern line.

[0019] As a further solution, the preset calibration loss function is constructed based on the following: missed items whose manual review conclusion is to block but whose review intensity requirement is lower than the corresponding intensity threshold, falsely blocked items whose manual review conclusion is to pass but whose review intensity requirement is not lower than the corresponding intensity threshold, and the total amount of review resources occupied. The missed items, falsely blocked items, and total amount of review resources occupied are weighted by the missed penalty coefficient, the falsely blocked penalty coefficient, and the review resource occupied coefficient, respectively, and include a regularization term for the deviation between the scene weight and the scene weight prior. The intensity threshold is solved by one-dimensional search and the scene weight is solved by gradient update iteratively; the interception rate parameter is automatically updated according to sliding window statistics; the calibration also includes updating the uncertainty penalty coefficient and modal weight, where the modal weight is the weight of each audit model in the comprehensive uncertainty; The review comments, reasons for rejection, and publication results are written into the audit chain record for the calculation of the preset calibration loss function. When the time-weighted average of the difference between the quantified value of the manual review conclusion and the corresponding review intensity requirement in the window remains deviated from zero within a preset time period, the time-weighted average is used as an additional term of the preset calibration loss function.

[0020] On the other hand, the present invention also provides an AI-based dynamic review process orchestration system based on multimodal risk scoring, used to execute the AI-based dynamic review process orchestration method as described in any of the preceding claims, comprising: The feature acquisition module is used to receive the object to be reviewed and its business context, perform multimodal analysis on text, image, video and audio data, and obtain dimensional risk scores for multiple risk dimensions. The risk calculation module is used to perform probability calculations on the category distribution output by the audit model to obtain the comprehensive uncertainty, to sum the risk scores of the dimensions according to the scenario weights to obtain the comprehensive risk score, and to calculate the audit intensity requirement based on the comprehensive risk score and the comprehensive uncertainty adjusted by the uncertainty penalty coefficient. The process orchestration module is used to configure interception rate parameters and resource consumption parameters for candidate review nodes, aggregate and calculate process strength based on the interception rate parameters of each review node, and automatically solve the target review process with the goal of minimizing the total review resource consumption among candidate review processes whose process strength is not less than the review strength requirement and meets the preset hard rules. The process execution module is used to automatically distribute audit tasks to the audit terminals corresponding to each audit node according to the target audit process. During the execution, it continuously acquires related monitoring data and recalculates the audit intensity requirements. When the preset adjustment conditions are met, it determines the remaining intensity requirements based on the process intensity of the executed audit nodes and automatically re-solves the remaining sub-processes. The calibration module is used to collect the conclusions of manual review, update the scene weights, intensity thresholds and interception rate parameters according to the preset calibration loss function, and feed them back to the risk calculation module, process orchestration module and process execution module. The audit log module is used to write the preset adjustment conditions, risk increments, and the audit nodes that are added or canceled and their risk basis into the audit chain record each time the target audit process is adjusted, and to write the audit opinion, rejection reason and release result into the audit chain record.

[0021] Compared with related technologies, the AI-based dynamic review process orchestration method and system based on risk scoring provided by this invention has the following advantages: 1. This invention combines the dimensional risk scores obtained from multimodal detection with model uncertainty to form the review intensity requirement. It automatically solves the review process with the constraint that the process intensity is not less than the review intensity requirement and with the goal of minimizing the total review resource consumption. This enables low-risk content to flow quickly and high-risk content to automatically obtain a review depth commensurate with its risk, reducing the overall system processing latency and the total review resource consumption. It solves the problem of mismatch between review depth and risk level in existing fixed processes.

[0022] 2. During process execution, this invention continuously acquires monitoring data and recalculates the review intensity requirements, automatically resolving the remaining sub-processes that have not yet been executed (including inserting review nodes, returning for re-review, and canceling high-resource nodes), and based on time extrapolation requirements, completes upgrades, triggers release gates, or performs forward-looking resource reductions before the risk actually crosses the threshold. This eliminates the lag of existing technologies that rely on manual reconfiguration of processes, and ensures that the process remains consistent as the risk status changes.

[0023] 3. This invention quantifies model output entropy, inter-model divergence, and missing dimension penalty into comprehensive uncertainty and incorporates them into the review intensity requirements. When model conclusions conflict, it automatically triggers manual review with accompanying evidence, thus making up for the deficiency in existing technologies where model misjudgments cannot be compensated by process mechanisms. At the same time, through the per-capability subspace constraint of dimensional requirement components and dimensional interception capability vectors, the review resources are allocated to the dimensions where the risks actually exist, avoiding the situation where nominal intensity is met but actual interception is insufficient.

[0024] 4. This invention uses a cost-sensitive preset calibration loss function to update scene weights, intensity thresholds, interception rate parameters, uncertainty penalty coefficients, and modal weights based on feedback from manual review conclusions, and absorbs system-level benchmark deviations, enabling the system to automatically correct itself after changes in business data distribution without manual parameter tuning; the audit chain records the adjustment conditions, risk increments, and the basis for adding or removing review nodes throughout the process, supporting the traceability of the generation and adjustment sources of the review process node by node, thus meeting the audit requirements. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the steps in the method provided in the embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the mapping of review intensity requirement levels provided in an embodiment of the present invention; Figure 3 The flowchart for process generation and re-arrangement provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the series-parallel model of audit strength provided in an embodiment of the present invention; Figure 5 The system overall architecture diagram provided for embodiments of the present invention; The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0029] Please see Figure 1 This invention provides an AI-based dynamic review process orchestration method based on risk scoring, executed by a computer system. The computer system is communicatively connected to multiple review models and multiple review terminals, and includes: S1: Receive the object to be reviewed and its business context, perform multimodal analysis on the text, image, video and audio data of the object to be reviewed, and obtain dimensional risk scores for multiple risk dimensions; S2: Perform probability calculation on the category distribution output by the audit model to obtain the comprehensive uncertainty; sum the dimensional risk scores according to the scenario weights to obtain the comprehensive risk score; calculate the audit intensity requirement based on the comprehensive risk score and the comprehensive uncertainty adjusted by the uncertainty penalty coefficient. S3: Configure interception rate parameters and resource consumption parameters for candidate review nodes, and aggregate and calculate the process strength based on the interception rate parameters of each review node in the candidate review process. The process strength increases monotonically with the increase of the number of review nodes participating in the aggregation and does not exceed 1. S4: Based on the constraint optimization algorithm, among the candidate review processes whose process intensity is not less than the review intensity requirement and meet the preset hard rules, the target review process is automatically solved with the goal of minimizing the total amount of review resources occupied, and the review tasks are automatically distributed to the review terminals corresponding to each review node according to the target review process. S5: During the execution of the target audit process, continuously acquire monitoring data associated with the object to be audited and recalculate the audit intensity requirement. When the preset adjustment conditions are met, determine the remaining intensity requirement based on the process intensity of the executed audit nodes, and automatically re-solve the remaining sub-processes that have not yet been executed in order to adjust the target audit process. S6: Collect the conclusions of manual review, and based on the deviation between the conclusions of manual review and the corresponding review intensity requirements, update the scene weight, intensity threshold and interception rate parameters according to the preset calibration loss function, and feed them back to S2 to S5.

[0030] The technical solution and innovations of the present invention will be described in detail below with reference to various embodiments: Example 1 like Figure 5 As shown, this embodiment provides an AI-based dynamic audit process orchestration system based on multimodal risk scoring. The system includes a feature acquisition module, a risk calculation module, a process orchestration module, a process execution module, a calibration module, and an audit record module. Each module is implemented by a computer system, which communicates with multiple audit models and multiple audit terminals.

[0031] The feature acquisition module receives the object to be reviewed and its business context, and calls the review model to perform multimodal analysis on the text, image, video, and audio data of the object to be reviewed, obtaining dimensional risk scores for multiple risk dimensions. The business context includes business object type (manuscript, resource, topic, instruction, activity, comment), tenant identifier, scenario category, publishing channel, dissemination parameters, submitter identity, associated activities or topics, and modification version number, etc.

[0032] The risk calculation module is used to perform probability calculations on the category distribution output by the audit model to obtain the comprehensive uncertainty U of the object to be audited. It also calculates the comprehensive risk score S by weighting the dimensional risk scores according to the scenario weights, and calculates the audit intensity requirement based on the comprehensive risk score and the comprehensive uncertainty adjusted by the uncertainty penalty coefficient.

[0033] The process orchestration module is used to configure interception rate parameters and resource consumption parameters for candidate review nodes. It aggregates and calculates the process strength based on the interception rate parameters of each review node in the candidate review process. In the candidate review process where the process strength is not less than the review strength requirement and meets the preset hard rules, it automatically solves the target review process with the goal of minimizing the total review resource consumption.

[0034] The process execution module is used to automatically distribute audit tasks to the audit terminals corresponding to each audit node according to the target audit process. During the execution, it continuously acquires monitoring data associated with the audit object and recalculates the audit intensity requirements. When the preset adjustment conditions are met, it determines the remaining intensity requirements based on the process intensity of the audit nodes that have been executed and automatically re-solves the remaining sub-processes that have not yet been executed.

[0035] The calibration module is used to collect the conclusions of manual review. Based on the deviation between the manual review conclusions and the corresponding review intensity requirements, it updates the scenario weight, intensity threshold and interception rate parameters according to the preset calibration loss function, and feeds the updated results back to the risk calculation module, process orchestration module and process execution module.

[0036] The audit log module is used to write the preset adjustment conditions, risk increments, and the audit nodes that are added or canceled and their risk basis into the audit chain record each time the target audit process is adjusted, and to write the audit opinion, rejection reason and release result into the audit chain record.

[0037] Based on the above system, the dynamic review process orchestration method of this embodiment includes the following steps.

[0038] In step S1, the object to be audited and its business context are received. Multimodal analysis is performed on the text, image, video, and audio data of the object to be audited to obtain dimensional risk scores for multiple risk dimensions. Risk dimensions include text content dimension, image and video dimension, audio dimension, person / organization dimension, factual credibility dimension, propagation and diffusion dimension, historical behavior dimension, tenant scenario dimension, and time event dimension. The original detection quantities for each dimension are heterogeneous and have different scales. In this embodiment, they are uniformly scaled to the [0,1] interval to obtain the dimensional risk score: logarithmic compression mapping is used for propagation magnitude, sigmoid function mapping is used for diffusion rate, lookup table mapping is used for subject sensitivity level, and piecewise linear mapping is used for historical violation rate.

[0039] In existing review systems, the detection results for text, images, and videos vary in format and scale. The system can only present these results side-by-side to reviewers and cannot convert them into quantities that can be used for process calculations, forcing process decisions to rely on manual intervention. This step maps the raw detection quantities of each risk dimension to a unified scale within the same numerical range, providing a unified computational basis for subsequent steps. This allows multi-source detection results to enter the same computational chain, thereby supporting the automation of process decisions.

[0040] In step S2, a probability calculation is performed on the category distribution output by the audit model to obtain the comprehensive uncertainty U. The dimensional risk scores are weighted and summed according to the scenario weights to obtain the comprehensive risk score S. Then, the audit intensity requirement is calculated based on the comprehensive risk score and the comprehensive uncertainty adjusted by the uncertainty penalty coefficient. The comprehensive risk score is calculated according to formula (1): In equation (1), For the k-th risk dimension, The scenario weight for this risk dimension in business context c is given, and the sum of the scenario weights is 1. The calculation of the overall uncertainty U is shown in Example 2. The audit intensity requirement is calculated according to formula (2): In equation (2), The uncertainty penalty coefficient, In this embodiment Take 0.5. From equation (2), it can be seen that the audit intensity requirement increases with the increase of the comprehensive risk score and with the increase of comprehensive uncertainty, and the degree of influence of comprehensive uncertainty decreases with the increase of the comprehensive risk score: when S approaches 1, the uncertainty coefficient ( When S approaches 0, the review intensity requirement is dominated by the comprehensive risk score; when S is low and U is high, the review intensity requirement is significantly increased, prompting the content to undergo manual confirmation. The multiplicative form of equation (2) used in this embodiment is only an example. Other function forms (such as weighted sum, exponential form, piecewise linear form) that satisfy the relationship of "increasing with increasing S, increasing with increasing U, and decreasing the influence of U as S increases" can be implemented equally.

[0041] This step incorporates two orthogonal pieces of information—risk magnitude and model credibility—into the decision-making process. Existing technologies typically rely solely on a single detection score for branching decisions, failing to differentiate between "how dangerous the content is" and "how reliable the model is," thus leaving confidence information idle and unrecoverable at the process level when the model misjudges. This step calculates these two separately and then synthesizes them into a single review intensity requirement. This allows information regarding insufficient model confidence or conflicting conclusions from multiple models to directly increase the review intensity requirement, compensating for model misjudgments through a process mechanism.

[0042] In step S3, interception rate parameters and resource consumption parameters are configured for candidate review nodes. The process strength is calculated by aggregating the interception rate parameters of each review node in the candidate review process. Candidate review nodes include machine review nodes, first-level manual review nodes, second-level review nodes, expert review nodes, leadership review nodes, external submission nodes, and countersigning groups, etc. Interception rate parameters... The probability of a review node v independently blocking inappropriate content, and its resource consumption parameters. This represents the resource consumption for a single audit corresponding to this audit node. The process intensity is calculated using formula (3): In equation (3), Let P be the set of review nodes that assume gatekeeping responsibilities in the candidate review process. Equation (3) means that the total miss probability of the process is the product of the miss probabilities of each review node, and the process strength is 1 minus the total miss probability. As can be seen from Equation (3), the process strength increases monotonically with the increase of the number of review nodes participating in the aggregation and does not exceed 1. The machine review node, as a pre-filter, is not configured with an interception rate parameter and does not participate in the aggregation calculation of the process strength; its insufficient reliability has been transmitted to the increase in the review strength requirement through comprehensive uncertainty, thus avoiding double calculation.

[0043] The reason for setting up this step is that the addition or removal of review nodes in the existing technology is configured based on experience, which cannot answer the calculation question of "whether the configured review is sufficient". This step abstracts the review nodes into interception rate parameters, and aggregates the process strength by multiplying the miss probability, so that "how strong the review is needed" and "how strong the review is actually provided by the process" can be compared on the same numerical scale, providing calculable constraints for the constraint solution in step S4.

[0044] In step S4, based on the constraint optimization algorithm, among the candidate review processes whose process intensity is not less than the review intensity requirement and meet the preset hard rules, the target review process is automatically solved with the goal of minimizing the total amount of review resources occupied, and the review tasks are automatically distributed to the review terminals corresponding to each review node according to the target review process. The total amount of review resources occupied is calculated according to formula (4): In equation (4), Total processing time of the process. This represents the timeliness weighting coefficient. Preset hard rules include review nodes that the review process must include or must not include, as well as tenant business rules. The target review process can be solved using the following greedy approach: First, select available review nodes according to the hard rules; then, add review nodes to the serial link sequentially from high to low according to the ratio of the interception rate parameter to the resource consumption parameter. Recalculate the process strength for each added node until the process strength is not less than the review strength requirement and the hard rules are satisfied. When the number of candidate review processes is large, the greedy approach can quickly obtain an approximately optimal solution that satisfies the constraints; when the number of candidate review processes is small, branch and bound or dynamic programming can be used to obtain the globally optimal solution. This embodiment uses a greedy approach only as an example.

[0045] This step addresses the problem of statically binding the review process to content risk in existing technologies. In the current fixed process, all content, regardless of risk level, follows the same path, resulting in low-risk content waiting in queues while high-risk content fails to receive in-depth review commensurate with its risk. This step automatically solves the process by constraining the process intensity to be no less than the review intensity requirement and aiming to minimize the total review resource consumption. This allows low-risk content to flow quickly, while high-risk content automatically receives the review depth matched to the requirements, achieving automatic matching of review resource consumption and risk status.

[0046] In step S5, during the execution of the target review process, monitoring data associated with the object to be reviewed is continuously acquired and the review intensity requirement is recalculated. When the preset adjustment conditions are met, the remaining intensity requirement is determined based on the process intensity of the executed review nodes, and the remaining sub-processes that have not yet been executed are automatically recalculated to adjust the target review process. The monitoring data includes dissemination channel data, associated event data, content modification data, etc. The preset adjustment conditions and the calculation of the remaining intensity requirement are described in Example 5.

[0047] This step addresses the problem of fixed processes in existing technologies failing to adapt to runtime changes. In existing systems, when content is modified during process execution, its dissemination scope expands, or the business environment changes, the generated processes cannot be adjusted accordingly, requiring manual reconfiguration, which results in significant lag. This step addresses this by having the system continuously acquire monitoring data, automatically recalculate the review intensity requirements, and resolve the remaining sub-processes. This ensures that the review process continuously matches the risk status during operation, eliminating the lag of manual reconfiguration, and making each adjustment traceable.

[0048] In step S6, the manual review conclusions are collected. Based on the deviation between the manual review conclusions and the corresponding review intensity requirements, the scene weights, intensity thresholds, and interception rate parameters are updated according to a preset calibration loss function, and then fed back to steps S2 to S5. The specific construction of the preset calibration loss function is shown in Example 8.

[0049] This step addresses the issue of review results not being fed back in existing technologies. In the current system, manual review conclusions, rejection reasons, and published results are not used to correct detection parameters and process thresholds, leading to gradual performance degradation as business data distribution changes. This step uses manual review conclusions as a calibration basis, updating scenario weights, intensity thresholds, and interception rate parameters, and feeding them back to the calculation and orchestration stages, forming a closed loop to ensure the system maintains decision-making effectiveness over long-term operation.

[0050] Figure 2 This illustrates the mapping method for review intensity requirements. Review intensity levels are mapped according to the intensity range into which the review intensity requirements fall: In this embodiment, Mapped to L0 level (automatic pass or sampling review). Mapped to Level L1 (Level 1 manual review). Mapped to Level L2 (multi-level manual or advanced review). This is mapped to Level L3 (expert, leadership, or external review). The thresholds of 0.40, 0.60, and 0.80 mentioned above are example values ​​and can be adjusted by the calibration module according to the tenant's business rules. The specific values ​​of the threshold parameters (incremental threshold, backoff threshold, uncertainty threshold, divergence threshold, lower limit of capability, preset capability increment, drift concern line, preset duration, etc.) involved in the various embodiments of this specification can be configured by the tenant according to business rules or automatically adjusted by the calibration module. The values ​​given in this specification are only examples and do not constitute a limitation of the present invention.

[0051] In a specific example, a routine event report consists of plain text with accompanying images. The risk scores for each dimension are as follows: text 0.1, image 0.1, people / organizations 0, factual credibility 0.2, dissemination and diffusion 0.3, and the scores for the other dimensions are relatively low. After weighting by scene weights, S=0.18; the confidence levels of each review model are high and the conclusions are consistent, with U=0.08. According to equation (2), we get If a content falls into Level L0, the system will automatically approve it and conduct a 10% sampling review, with no human intervention throughout the process, reducing processing time from hours to seconds. This example demonstrates that low-risk content with a reliable model does not need to wait in line for manual review; review resources are automatically allocated to high-risk content.

[0052] Example 2 This embodiment further explains the calculation of the comprehensive uncertainty U in step S2. The comprehensive uncertainty U is obtained by weighted combination of model output uncertainty, inter-model divergence, and missing dimension penalty term, and is calculated according to equation (5): In equation (5), , , Let be the combination coefficient, and .

[0053] The uncertainty of the model output is obtained by weighting the normalized entropy of the category distribution of each audit model output according to the modal weights, and is calculated according to equation (6): In equation (6), The category distribution output by the m-th review model. The normalized entropy of this category's distribution. These are modal weights. Higher entropy indicates that the model hesitates more between categories.

[0054] The inter-model divergence is the mean of the differences between the category distributions output by any two audit models. In this embodiment, the divergence is measured using the Jensen-Shannon divergence (JS divergence), calculated according to equation (7): In equation (7), M represents the total number of audit models. Output the JS divergence between the category distributions of the i-th and j-th review models. The difference can also be achieved using other distance metrics between distributions, such as KL divergence or maximum mean difference; JS divergence is only an example. This metric increases significantly when multiple review models are deployed in the same modality (e.g., proprietary models, general large models, third-party review services), or when review conclusions from different modalities conflict (e.g., low-risk text but high-risk visuals).

[0055] The penalty for missing dimensions is the sum of the scenario weights corresponding to the missing risk dimensions, calculated according to equation (8): In equation (8), This is a set of risk dimensions for the object to be reviewed that are missing. When the object to be reviewed lacks a certain modality (e.g., a plain text document without images), the scenario weight for the missing dimension is included. Meanwhile, the scene weights of the remaining available dimensions are renormalized proportionally. The effect is that "image cannot be parsed" itself increases the overall uncertainty, prompting manual review, rather than being ignored as risk-free.

[0056] The dimensional risk scores for image and video risk dimensions are obtained by parallel aggregation of the unit risk scores of multiple image units, calculated according to equation (9): In equation (9), Let be the unit risk score of the j-th frame unit, which is the detection risk score of a single image for images and the detection risk score of each frame after time window extraction for videos. Equation (9) means that any frame unit that violates the rules is considered a complete hit, and the aggregation result increases monotonically with the number of frame units and does not exceed 1.

[0057] The risk score for the audio risk dimension is calculated from the transcribed text of the audio data using the text risk dimension. When the audio contains specific sensitive voiceprints, the risk score for that dimension is directly determined according to the level of the sensitive voiceprint database.

[0058] The uncertainty calculation in this embodiment is used to characterize the "uncertain" state of the review model. If only the detection score is used and this state is ignored, then when the object to be reviewed is missing a certain modality, the model output is hesitant, or different model conclusions conflict, the system has no way to express the degree of information incompleteness and may treat missing information as risk-free. Through the combination of the above three items, information incompleteness, model hesitation, and conclusion conflict are all quantified as comprehensive uncertainty and incorporated into the review intensity requirements, prompting the corresponding content to enter the manual confirmation stage, achieving the effect of "better to review more than to miss".

[0059] Example 3 This embodiment is in Figure 2 Further explanation is provided based on the mapping shown. Solving the target review process also includes three steps: level mapping, hard label constraints, and disagreement review.

[0060] First, the review intensity level is mapped according to the intensity range that the review intensity requirement falls into, as described in Example 1. This level mapping automatically classifies the review depth according to the risk level, preventing all content from being treated equally.

[0061] Secondly, when the subject of review is identified by a key risk label, the review intensity level is constrained to be no lower than the minimum review intensity level corresponding to that label. Key risk labels include “important person,” “major emergency,” “unverified medical statement,” and “suspected deepfake.” Key risk labels do not participate in the weighted summation of the comprehensive risk score, but serve as a hard constraint for the process solution, preventing high-weight dimensions from obscuring key facts. For example, if an interview article is identified by the key risk label “important person,” and the minimum review intensity level corresponding to this label is L2, then regardless of which intensity range the review intensity requirement calculated according to formula (2) falls into, the review intensity level of the target review process must not be lower than L2. This hard constraint ensures that specific risk content receives mandatory minimum review protection and is not diluted by weighted averaging.

[0062] Furthermore, when the overall uncertainty exceeds the uncertainty threshold, or the difference between the category distributions output by any two review models exceeds the divergence threshold, a manual review node is added as a mandatory review node to the target review process. The conclusions, confidence levels, and matched text snippets, video frames, or source annotations of each review model are sent along with the review task to the review terminal corresponding to the manual review node. The matched text snippets are the prohibited words and phrases highlighted by the text review model, the video frames are the suspicious frames marked by the video review model, and the source annotations are the reference locations of contradictory evidence in the factual credibility dimension. This step provides a fallback for situations where model conclusions are unreliable, submitting conflicts between models along with the evidence to human reviewers to prevent conflicts from being masked by simple voting.

[0063] In a specific example, a popular science article was detected by two text moderation models: Model A output a class distribution of [0.05, 0.95] (low risk), while Model B output a class distribution of [0.60, 0.40] (suspected violation). The JS divergence between the two models was 0.58, exceeding the divergence threshold. The divergence between models increased significantly, and the overall uncertainty U rose to 0.72. Calculated according to equation (2) The system progresses from L0 to L1 and is approaching the L2 threshold. A manual review node is then inserted, sending the conclusions, divergent segments, and source annotations of both models along with the review task. After manual evaluation determines it to be a reasonable academic discussion, the sample, along with the manual review conclusion, enters the calibration pool, and the modal weights of model B on this topic are reduced (see Example 8).

[0064] Example 4 This embodiment further explains steps S3 and S4. The relationship between the process strength aggregation of serial review nodes and the calculation of the equivalent interception rate of the countersigning group is as follows: Figure 4 As shown.

[0065] Regarding the setting and updating of the interception rate parameter, the interception rate parameter represents the probability that an audit node independently intercepts inappropriate content. Its initial value is preset based on professional division of labor and historical audit performance; for example, 0.60 for first-level manual audit nodes, 0.55 for second-level audit nodes, 0.75 for senior audit nodes, and 0.90 for supervisor audit nodes. After running, the interception rate parameter is automatically updated according to the ratio of the number of pieces of content actually intercepted and confirmed as inappropriate within the sliding window to the total number of pieces of content that enters the audit node and is confirmed as inappropriate. In equation (10), This represents the number of pieces of content that were actually intercepted by the review node within the sliding window and confirmed as violations by manual review. This represents the total number of content items that enter this review node within the sliding window and are confirmed as violations by manual review. This update method automatically adjusts the process strength model based on the long-term actual performance of review nodes, avoiding long-term overestimation or underestimation of review capabilities based on initial experience values.

[0066] When there is a joint review group with multiple reviewers conducting parallel reviews, the joint review group participates in the process strength calculation as an equivalent review node, and its equivalent interception rate is calculated according to formula (11): In equation (11), Let be the interception rate parameter for each auditor in the joint signing group. As can be seen from equation (11), the equivalent interception rate is higher than the interception rate parameter of any auditor in the joint signing group, and increases with the number of joint signers, but does not exceed 1. This equivalent processing enables the contribution of parallel review to be quantified and included in the process intensity, avoiding the joint signing from becoming a mere formality of repetitive nodes.

[0067] The preset hard rules include review nodes that must be included or prohibited, as well as tenant business rules. For example, content involving important figures must include a high-level review node, and tenants stipulate that major news must go through an external review node. The calculation of the total review resource usage also includes a latency term, as shown in equation (4).

[0068] In a specific example, an interview article hits the key risk tag "important person," and the tag's hard constraint requires a review intensity level of no less than L2. Calculated values ​​S=0.62, U=0.15, and R=0.65, it falls into the L2 level. There are three feasible candidate review processes: Process A (machine review, first-level manual review, second-level review), Process B (machine review, first-level manual review, senior review), and Process C (machine review, first-level manual review, second-level review, leadership review). Calculated according to formula (3), the process intensity of Process A is... ,satisfy However, the review intensity level of process A is L1, which does not meet the hard constraint of the label; the process intensity of process B is... Process C satisfies both strength and level constraints; process C has higher process strength but includes a leader node, resulting in greater resource consumption. The system selects process B, which has the lowest total resource consumption among all processes satisfying all constraints, as the target review process. This example demonstrates that solving for the target review process involves minimizing resources under the combined influence of hard and strength constraints, rather than simply stacking nodes.

[0069] Example 5 This embodiment further explains step S5, which can be combined with... Figure 3 understand.

[0070] The preset adjustment conditions include any of the following: recalculated review intensity requirements. The difference between the audit intensity requirement at the time of arrangement and the incremental threshold (the incremental threshold is 0.15 in this embodiment); the audit intensity level is out of tier; a new key risk label is hit; or the object to be audited is resubmitted with a new version.

[0071] When the preset adjustment conditions are met, the remaining strength requirements of the remaining sub-processes decrease as the process strength of the executed audit nodes increases. In this embodiment, the remaining strength requirements are determined according to formula (12): In equation (12), To recalculate the required level of review intensity, The process strength of the audited nodes is defined as follows: The derivation process of equation (12) is as follows: The overall strength must meet the following requirements. Thus, the solution is found. The lower limit. Within the remaining feasible space, the system re-solves the remaining sub-processes as described in Example 4. Executable actions include inserting higher-level review nodes, upgrading single-person review to countersigning, and adding external submission nodes. Due to the remaining strength requirement... The larger the contribution of the already executed part, the less effort is needed to supplement the remaining sub-processes, thus avoiding the need to repeatedly add review resources to content that has already been fully reviewed.

[0072] In cases where the case is returned for retrial, when When the difference between the audit intensity requirement at the time of arrangement and the actual audit status exceeds the return threshold, it indicates that the risk increment is so large that the already executed audit conclusion is unreliable. The system will return the object to be audited to the previous audit node for re-audit, along with risk evidence from the recalculation. Risk evidence includes detailed dimensional risk scores from the recalculation, changes in monitoring data that triggered the adjustment, and newly added key risk tags. Returning the object for re-audit ensures that the already executed audit conclusion remains consistent with the current risk status.

[0073] Regarding the reduction in features, when When a decrease leads to a downgrade in the review intensity level, the system cancels review nodes that have not yet been started and whose resource usage parameters are higher than the preset resource threshold, releasing the already occupied review resources so that the review resource usage shrinks in a timely manner as the risk decreases.

[0074] The system also maintains an audit chain record. Each time the target audit process is adjusted, the preset adjustment conditions, risk increments, and the added or removed audit nodes and their risk basis are written into the audit chain record. This is used to trace the source of the target audit process adjustment by audit node. The structured entries in the audit chain record include: adjustment timestamp, type of preset adjustment condition triggered, etc. The difference between the audit intensity requirements during the arrangement, the added or removed audit node identifiers and their interception rate parameters, the process differences before and after the adjustment, and the citation of risk evidence.

[0075] In a specific example, a news article concerning people's livelihood is being reviewed according to the L1 level process, and is currently undergoing a Level 1 manual review. Prior to its release, its distribution channels expanded from WeChat official accounts to online distribution across the entire network. The risk score for the dissemination and diffusion dimension rose from 0.2 to 0.7, and it entered a sensitive window for sudden events, requiring recalculation. , The difference between the strength requirement approved during the compilation process and the actual strength requirement is 0.32, exceeding the incremental threshold of 0.15. The remaining strength requirement is obtained using equation (12). The system inserts a two-person expert countersigning group into the remaining sub-processes (equivalent interception rate). Meanwhile, due to a significant increase in risk, reviewed but unpublished content, along with evidence of risks related to "channel expansion + sensitive windows," was returned to the previous review stage for supplementary review. The entire process was completed automatically by the system, requiring no manual reconfiguration of the workflow, and all adjustments were recorded in the audit chain.

[0076] Example 6 This embodiment further explains the dimensional decomposition and capability matching of review intensity requirements. The dimensional processing in this embodiment addresses the shortcomings of scalar intensity constraints: scalar constraints only ensure that the total amount of review intensity is sufficient, but cannot guarantee the correct direction of review. The same review intensity requirement may come from completely different risk dimensions. If the same review resources are allocated indiscriminately, a situation may arise where the nominal intensity is met, but the actual interception effect is insufficient. By placing the requirement and capability in the same capability subspace for measurement and constraining each subspace, review resources are guided to the dimension where the risk actually resides, achieving "review in the right direction."

[0077] Regarding the decomposition of requirements, the audit intensity requirement is decomposed into dimensional audit intensity requirement components according to the risk dimension, and each requirement component is calculated according to formula (13): In equation (13), For the dimension of risk score, As scene weight, This represents dimensional uncertainty. As shown in equation (13), each demand component increases with increasing dimensional risk score and dimensional uncertainty, while the impact of dimensional uncertainty decreases with increasing dimensional risk score; the sum of all demand components is the audit intensity demand R, i.e. The scalar form of equation (2) can be regarded as an aggregate approximation of the vector form of equation (13) under the condition of amortization of uncertainty.

[0078] Dimensional uncertainty It consists of the following three parts: the entropy contribution of the category group mapped to the risk dimension in the category distribution output by each audit model, the component mapped to the risk dimension in the inter-model divergence, and the missing dimension penalty term of the risk dimension. Among them, the category group is the set of categories belonging to the risk dimension in the category distribution; the entropy contribution and the divergence component are obtained by summing the normalized entropy of Equation (6) and the pairwise JS divergence of Equation (7) according to the dimension of the violation category.

[0079] In terms of capability configuration, dimensional interception capability vectors are configured for candidate review nodes. Capability subspaces are formed by aggregating multiple risk dimensions according to professional division of labor. In this embodiment, nine risk dimensions are aggregated into four capability subspaces: text content (text content, people / organizations), visual (images / videos), factual source (factual credibility, time-based events), and subject / business (propagation / diffusion, historical behavior, tenant scenarios, audio). Tenants can customize subspace divisions according to their actual division of labor. Dimensional interception capability vectors ,in To assess the dimensional interception capability of node v in the capability subspace c, its initial value is preset by the professional division table (for example, financial editors have high dimensional interception capability in the factual information source class subspace and medium capability in the text content class subspace), and is automatically updated after running as described later in this embodiment.

[0080] In terms of constraint solving, the capacity subspace is used as the constraint unit, and the sum of all demand components mapped to the same capacity subspace is taken as the subspace strength demand, i.e. The subspace process strength of each capability subspace is calculated based on the dimensional interception capabilities of each review node, and is calculated according to formula (14): The target review process is solved with the constraint that the subspace process intensity of each active subspace is not less than the corresponding subspace intensity requirement. An active subspace is defined as a subspace whose subspace intensity requirement exceeds the dimension participation threshold (in this embodiment, the dimension participation threshold is...). (Take 0.08), subspaces below this threshold do not constitute constraints, thus avoiding process expansion due to noise dimensions.

[0081] When solving the target review process, the intensity gap of each active subspace is calculated according to formula (15): Furthermore, candidate review nodes with higher interception capabilities and lower resource consumption parameters on unclosed strength gaps are prioritized for inclusion in the target review process. This embodiment uses gap-weighted cost-effectiveness. The selection criteria are calculated according to formula (16): In equation (16), These are the resource usage parameters for candidate review nodes. Each selection... The largest candidate review node is added to the target review process, and the strength gaps of each active subspace are updated until all strength gaps are closed. The improvement of Equation (16) over the scalar cost-effectiveness lies in the fact that the value of node selection changes from "total interception amount" to "the amount intercepted at the current gap", which is implemented as a vector dot product.

[0082] When a joint approval group exists, a combination of complementary capabilities from each active subspace is selected from the candidate review nodes based on joint capabilities. The joint capabilities are calculated according to formula (17): For example, the roles of "political editor + data verifier" complement each other in the two subspaces of text content and factual source. If unclosed strength gaps still exist after traversing available candidate review nodes, an external review node is automatically inserted. In the system, an external review node is represented as a review node with high-dimensional interception capabilities and high resource consumption parameters; its trigger condition is that the internal capability gap cannot be closed.

[0083] During execution, if the current audit node's dimensional interception capability in a certain active subspace is lower than the lower limit of capability, and there is an idle candidate audit node whose dimensional interception capability in that subspace is higher than the current audit node by a preset capability increment, then the audit will be automatically transferred to that idle candidate audit node, and the transfer basis (capability comparison) will be written into the audit chain record.

[0084] After execution, the dimensional interception capability of each review node is automatically updated subspace by subspace based on the ratio of the number of actually intercepted and confirmed violations within the sliding window to the total number of violations entering that review node and confirmed to be violations in the corresponding capability subspace. If a capability subspace fails to process content within a predetermined number of times within a predetermined time window, even if the subspace process strength meets the constraints, it indicates that the capability estimate for that subspace is too high. The dimensional interception capability of each review node within that subspace is then adjusted by a predetermined shrinkage factor. Lower (in this embodiment) (Take 0.9), so that the dimensional constraints can be restored to correspond with the actual interception capabilities.

[0085] In a specific example, Article A is a political commentary, with risks concentrated in the text content subspace (subspace strength requirement 0.62); Article B is a financial report analysis, with risks concentrated in the factual source subspace (subspace strength requirement 0.58). Both articles have a review strength requirement of 0.65. Without dimension matching, both articles would receive the same review process strength. After dimension matching, Article A is covered by a political editor (text content dimension interception capability 0.82) plus leadership oversight, while Article B is covered by a financial expert (factual source dimension interception capability 0.80) plus data verification and countersigning. Article B does not require leadership involvement, reducing the total review resource consumption by approximately 30%, while the subspace process strength of both articles in their respective risk subspaces actually increases from 0.75 to 0.85 / 0.83.

[0086] In another specific example, the calibration module found that after the factual source interception capability of an editor was corrected from 0.70 to 0.55, the subspace process strength of the financial review process independently undertaken by the editor decreased from 0.70 to 0.55, which is lower than the subspace strength requirement of 0.58. The system automatically added a second verification node to the process without requiring any manual modification of the process template.

[0087] Example 7 This embodiment further explains the situation where the judgment of preset adjustment conditions is based on time extrapolation requirements. The time extrapolation in this embodiment is designed to address the lag problem triggered after the fact: the existing technology only adjusts the process after the risk actually crosses the threshold. At this time, the remaining processing time is often insufficient to accommodate the new review process, resulting in timeouts or releases with defects. By incorporating the drift rate, the remaining processing time and the deviation from the system-level benchmark into the decision, the adjustment can be made before the risk actually crosses the threshold, leaving time margin for the new review process.

[0088] The system maintains a risk sampling sequence and estimates the risk drift rate based on changes in review intensity requirements at adjacent sampling points. Sources of the risk sampling sequence include: content modifications (version re-releases), changes in distribution channels or parameters, event calendar updates, and periodic scanning of content in transit by the execution engine. The risk drift rate is estimated using the exponentially weighted slope of equation (18): In equation (18), , The review intensity requirements for two adjacent sampling points. , The sampling time corresponding to the two adjacent sampling time points. The smoothing coefficient is set to (0,1), and is 0.3 in this embodiment. To suppress single-sample noise, the drift term is enabled only when the number of sampling points is not less than 2, and no update is performed when the interval between adjacent sampling points is less than the minimum interval (60 seconds in this embodiment).

[0089] The system also updates the system-level benchmark deviation based on the time-weighted average of the difference between the quantitative value of the manual review conclusion and the corresponding review intensity requirement, calculated using the exponential weighted average of formula (19): In equation (19), This is the quantitative value of the manual review conclusion (1 for a blocked result and 0 for a passed result). This represents the required level of review intensity for the manual review conclusion. The smoothing coefficient is set to (0,1). It is updated once for each manual review conclusion. The system-level benchmark deviation is included in the time extrapolation requirement only when the system-level benchmark deviation exceeds the preset control band; within the band, it is zeroed out (in this embodiment, the preset control band is...). This ensures that daily random fluctuations do not affect decision-making, and only persistent and systematic biases during the event period are shifted to the baseline. The role of system-level baseline bias is to absorb the overall distortion of model scores during the event period, avoiding the situation where individual corrective measures cannot keep up with the overall drift.

[0090] Based on this, the time extrapolation requirement is calculated using the product of the current sampling time's audit intensity requirement, risk drift rate, and remaining processing time (after truncation) and the sum of the system-level baseline deviation, as well as the result of the truncated product. The calculation follows formula (20): In equation (20), Based on the review intensity requirements at the current sampling point, The remaining processing time is the sum of the delays of the remaining review nodes in the current target review process; for news scenarios, the publication deadline can also be used. To the upper limit of projection A cutoff function bounded by a boundary, in this embodiment Take 0.25, This is a system-level baseline deviation that accounts for time extrapolation requirements. The physical meaning of time extrapolation requirements is not "how dangerous it is now," but rather "how high the expected risk will be by the time of review or release."

[0091] Furthermore, the recalculated review intensity requirement is replaced with the time extrapolation requirement, and the review intensity level determination and preset adjustment conditions are judged. When the time extrapolation requirement exceeds the intensity threshold of the corresponding interval of the current review intensity level, the review process is upgraded in advance. The upgrade occurs before the actual grade change, and the latency required for the new review node is included in the remaining path in advance. When the remaining processing time is less than the latency required for the new review node, the release of the object to be reviewed is suspended until the reinforcement review is completed (release gate), and this action and its basis are written into the audit chain record. When the time extrapolation requirement is lower than the lower limit of the current review intensity level and continues for a preset duration (10 minutes in this embodiment), the review nodes that have not yet been started and whose resource consumption parameters are higher than the preset resource threshold are canceled. In addition, the risk sampling cycle is shortened for the object to be reviewed whose absolute value of risk drift rate exceeds the drift concern line, and low-frequency scanning is maintained for objects with drift close to zero, so that scanning resources are tilted towards content that is actually changing.

[0092] In a specific example, a press release about a sudden incident enters the L1 level review queue, and the current review intensity requirement is... (Not reaching the L2 threshold of 0.60), but the risk escalates due to the forwarding and dissemination of the article. Minutes, remaining processing time Minutes. Calculate according to formula (20). The system pre-inserts expert review nodes and adjusts subsequent paths. The actual review intensity requirement only crosses 0.60 after 15 minutes: if the upgrade were only performed after the threshold is crossed using the mechanism without time extrapolation, the remaining processing time would be only 5 minutes, while adding an expert review would take 12 minutes, which would only trigger the release gate or result in a release with defects; in this embodiment, the upgrade action saves 15 minutes, and the entire process does not exceed the time limit.

[0093] In another specific example, following a public health emergency in a certain area, new terminology and new entities caused the review model to systematically under-score all content in transit. Within three days, the system-level baseline deviation rose to +0.12, exceeding the preset control zone. , The changes took effect automatically, with the review intensity benchmark for all content in transit across the system uniformly increased, resulting in a large number of drafts being reinforced in advance. Five days after the event ended, the system-level benchmark deviation dropped back to 0.03, entering the in-band zeroing phase, and the benchmark automatically reverted, requiring no manual intervention throughout the process.

[0094] In another specific example, after a sensitive time window was lifted, the risk sampling sequence of a batch of queued manuscripts turned down, the risk drift rate became negative, the time extrapolation demand was lower than the current review intensity level limit and lasted for more than 10 minutes, the system canceled the leadership review nodes that had not yet been started, saving about 2.5 person-hours of senior review manpower per day.

[0095] Example 8 This embodiment further explains step S6. The closed-loop calibration in this embodiment continuously adjusts the threshold, weight, and interception capability according to the actual business performance. Without this step, the risk score and process threshold will remain unchanged for a long time, gradually becoming distorted after changes in business data distribution. Through a cost-sensitive calibration loss function and alternating iterative solutions, the system can automatically converge to a configuration suitable for the current business among the three factors of missed releases, false blocks, and resource consumption. Moreover, the data used for calibration and the audit data are from the same source, and the entire process is traceable.

[0096] The preset calibration loss function is constructed based on the following: missed items whose manual review conclusion is blocking but whose review intensity requirement is lower than the corresponding intensity threshold; falsely blocked items whose manual review conclusion is passing but whose review intensity requirement is not lower than the corresponding intensity threshold; and the total review resource usage. Missed items, falsely blocked items, and total review resource usage are weighted by the missed penalty coefficient, the falsely blocked penalty coefficient, and the review resource usage coefficient, respectively, and include a regularization term for the deviation between the scene weight and the prior scene weight. This embodiment is constructed according to formula (21): In equation (21), This is the quantitative value of the manual review conclusion (1 for a blocked result and 0 for a passed result). This represents the required level of review intensity for the manual review conclusion. This is the intensity threshold for the corresponding gear level (each gear level can be set separately). For indicator functions, This is the penalty coefficient for missed releases. This is the penalty coefficient for wrongful blocking. To review resource usage coefficients, This refers to the total amount of review resources consumed in the review process for object j to be reviewed. As scene weight, As a priori for scene weights, This is a regularization term used to prevent scene weights from deviating excessively from the tenant's prior rules. Since the penalty coefficients for missed releases and false blockings are configured by the tenant according to the business consequences (the cost of missing political news is much higher than that of entertainment content), different tenants and scenes will naturally converge to different intensity threshold systems, achieving scene-adaptive thresholds.

[0097] In terms of solution method, the intensity threshold is solved by one-dimensional search, and the scene weight is solved by gradient update. The one-dimensional search and gradient update are executed alternately and iteratively. The interception rate parameter is automatically updated according to the sliding window statistics, as shown in Equation (10); the dimensional interception capability is updated subspace by subspace according to the same statistical method as Equation (10). Calibration also includes updating the uncertainty penalty coefficient and modal weights. The modal weights are the weights of each audit model in the comprehensive uncertainty: when the conclusion of an audit model on a certain topic is repeatedly rejected by the manual audit conclusion, its modal weights are adjusted accordingly. Audit opinions, rejection reasons and release results are written into the audit chain record for the calculation of the preset calibration loss function, so that the data used for calibration is from the same source as the audit data.

[0098] Furthermore, when the time-weighted average of the difference between the quantified value of the manual review conclusion within the window and the corresponding review intensity requirement remains deviating from zero within a preset time period, this average is used as an additional term in the preset calibration loss function. This means that if the manual review conclusion continuously deviates from the model's calculated value within a certain period, it indicates a structural bias in the weights or thresholds. During periodic calibration, this window's average is incorporated into the calibration target to ensure that the weight update direction aligns with the online benchmark shift direction.

[0099] In a specific example, after a tenant ran the system for a quarter, the calibration module found that L0-level missed releases were concentrated in the range [0.32, 0.40); the scenario weight of the text content dimension was increased by 8% due to multiple missed releases; and the interception rate parameter of a certain review node was corrected from the initial value of 0.70 to 0.82. The missed release rate decreased by 37% in the following quarter, while the automatic approval rate only decreased by 6%, and the total review resource usage remained basically the same. This example demonstrates that closed-loop calibration continuously reduces the missed release rate without significantly increasing review resource usage, and the system can adapt to business changes without manual parameter tuning.

[0100] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for orchestrating a dynamic review process based on risk scoring using artificial intelligence, executed by a computer system, wherein the computer system is communicatively connected to multiple review models and multiple review terminals, characterized in that... include: S1: Receive the object to be reviewed and its business context, perform multimodal analysis on the text, image, video and audio data of the object to be reviewed, and obtain dimensional risk scores for multiple risk dimensions; S2: Perform probability calculation on the category distribution output by the audit model to obtain the comprehensive uncertainty; sum the dimensional risk scores according to the scenario weights to obtain the comprehensive risk score; calculate the audit intensity requirement based on the comprehensive risk score and the comprehensive uncertainty adjusted by the uncertainty penalty coefficient. S3: Configure interception rate parameters and resource consumption parameters for candidate review nodes, and aggregate and calculate the process strength based on the interception rate parameters of each review node in the candidate review process. The process strength increases monotonically with the increase of the number of review nodes participating in the aggregation and does not exceed 1. S4: Based on the constraint optimization algorithm, among the candidate review processes whose process intensity is not less than the review intensity requirement and meet the preset hard rules, the target review process is automatically solved with the goal of minimizing the total amount of review resources occupied, and the review tasks are automatically distributed to the review terminals corresponding to each review node according to the target review process. S5: During the execution of the target audit process, continuously acquire monitoring data associated with the object to be audited and recalculate the audit intensity requirement. When the preset adjustment conditions are met, determine the remaining intensity requirement based on the process intensity of the executed audit nodes, and automatically re-solve the remaining sub-processes that have not yet been executed in order to adjust the target audit process. S6: Collect the conclusions of manual review, and based on the deviation between the conclusions of manual review and the corresponding review intensity requirements, update the scene weight, intensity threshold and interception rate parameters according to the preset calibration loss function, and feed them back to S2 to S5.

2. The method according to claim 1, characterized in that, The overall uncertainty is obtained by weighted combination of model output uncertainty, inter-model divergence, and missing dimension penalty term; wherein, the model output uncertainty is obtained by weighted summation of the normalized entropy of the category distribution output by each review model according to modal weights, the inter-model divergence is the mean of the difference between the category distribution output by any two review models, and the missing dimension penalty term is the sum of the scenario weights corresponding to the missing risk dimension.

3. The method according to claim 1, characterized in that, Solving the target review process also includes: The review intensity level is mapped according to the intensity range into which the review intensity requirement falls; when a key risk label is hit, the review intensity level is constrained to be no lower than the minimum review intensity level corresponding to that label; When the overall uncertainty exceeds the uncertainty threshold, or the difference between the category distributions output by any two audit models exceeds the divergence threshold, the manual review node is added to the target audit process as a mandatory audit node, and the conclusions, confidence levels, and hit text fragments, video frames, or source annotations of each audit model are sent to the audit terminal corresponding to the manual review node along with the audit task.

4. The method according to claim 1, characterized in that, The interception rate parameter is the probability that an audit node independently intercepts illegal content. Its initial value is preset according to professional division of labor and historical audit performance, and it is automatically updated after running according to the ratio of the number of illegal content actually intercepted and confirmed in the sliding window to the total number of illegal content entering the audit node and confirmed. When there is a joint review group with multiple reviewers conducting parallel reviews, the joint review group participates in the process strength calculation as an equivalent review node, and its equivalent interception rate is higher than the interception rate parameter of any reviewer in the joint review group; the preset hard rules include review nodes and tenant business rules that must be included or prohibited from being included; the calculation of the total amount of review resources also includes a latency item.

5. The method according to claim 3, characterized in that, The preset adjustment conditions include: recalculated review intensity requirements. The difference between the audit intensity requirement at the time of arrangement and the incremental threshold is exceeded, the audit intensity level is crossed, the newly added key risk label is hit, or the audit object is resubmitted with a new version; When the preset adjustment conditions are met, the remaining strength requirements that the remaining sub-processes need to meet decrease as the process strength of the executed audit nodes increases, and the remaining sub-processes are automatically re-solved accordingly. when If the difference between the audit intensity requirement during the arrangement exceeds the return threshold, the object to be audited will be returned to the previous audit node for re-audit with attached risk evidence. when When a decrease leads to a downgrade in the review intensity level, cancel any review nodes that have not yet been started and whose resource usage parameters are higher than the preset resource threshold. Each time the target audit process is adjusted, the preset adjustment conditions, risk increments, and the audit nodes that are added or canceled, along with their risk basis, are written into the audit chain record. This is used to trace back the source of the adjustment to the target audit process by audit node.

6. The method according to claim 2, characterized in that, The audit intensity requirement is decomposed into dimensional audit intensity requirement components according to the risk dimension. Each requirement component increases with the increase of the dimension risk score and dimension uncertainty, and the influence of dimension uncertainty decreases with the increase of the dimension risk score. The sum of each requirement component is the audit intensity requirement. The dimensional uncertainty consists of the entropy contribution of the category group mapped to the risk dimension in the category distribution output by each audit model, the component mapped to the risk dimension in the inter-model divergence, and the missing dimension penalty term of the risk dimension; a dimensional interception capability vector is configured for the candidate audit node; Using the capability subspace as the constraint unit, the sum of each requirement component mapped to the same capability subspace is taken as the subspace strength requirement. The subspace process strength on each capability subspace is calculated by aggregating the interception capabilities of each review node. The target review process is solved with the constraint that the subspace process strength of each active subspace is not less than the corresponding subspace strength requirement. The capability subspace is formed by aggregating multiple risk dimensions according to professional division of labor, and the active subspace is the subspace whose intensity requirement exceeds the dimension participation threshold.

7. The method according to claim 6, characterized in that, When solving the target review process, the strength gap of each active subspace is calculated, and candidate review nodes with higher dimensional interception capabilities and lower resource consumption parameters on unclosed strength gaps are preferentially selected to join the target review process until all strength gaps are closed. When a countersigning group exists, select a combination of complementary capabilities from each active subspace; if there is still an unclosed strength gap after traversing the available candidate review nodes, automatically insert an external submission node. During execution, if the current review node's dimensional interception capability in a certain active subspace is lower than the lower limit of its capability, and there is an idle candidate review node whose dimensional interception capability in that subspace is higher than the current review node's preset capability increment, then the review will be automatically transferred to that idle candidate review node. After execution, the dimensional interception capability of each review node is automatically updated subspace by subspace according to the ratio of the number of content that is actually intercepted and confirmed to be in violation within the sliding window to the total number of content that enters the review node and is confirmed to be in violation in the corresponding capability subspace. When a capability subspace fails to meet the constraints but still misses a process if it exceeds a preset number of times within a preset time window, the dimensional interception capability of each audit node in the subspace is reduced by a preset shrinkage factor.

8. The method according to claim 5, characterized in that, The determination of the preset adjustment conditions is based on the time extrapolation requirement, and the calculation of the time extrapolation requirement includes: Maintain the risk sampling sequence and estimate the risk drift rate based on the changes in the review intensity requirements at adjacent sampling time points; update the system-level benchmark deviation based on the time-weighted average of the difference between the quantified value of the manual review conclusion and the corresponding review intensity requirements, and only include the time extrapolation requirement when the system-level benchmark deviation exceeds the preset control band; The time extrapolation requirement is the sum of the audit intensity requirement at the current sampling time, the product of the risk drift rate and the remaining processing time after truncation, and the system-level benchmark deviation. The audit intensity level determination and the judgment of the preset adjustment conditions are performed by replacing the recalculated audit intensity requirement with the time extrapolation requirement. When the time extrapolation requirement exceeds the intensity threshold of the current review intensity level, the review process is upgraded in advance; when the remaining processing time is less than the delay required for adding a new review node, the release of the object to be reviewed is suspended until the supplementary review is completed. When the time extrapolation requirement is lower than the current review intensity level lower limit and continues for a preset duration, cancel the review node whose resource usage parameter is higher than the preset resource threshold that has not yet been started; shorten the risk sampling cycle for the pending review object whose absolute value of risk drift rate exceeds the drift concern line.

9. The method according to claim 1, characterized in that, The preset calibration loss function is constructed based on the following: missed items whose manual review conclusion is to block but whose review intensity requirement is lower than the corresponding intensity threshold, falsely blocked items whose manual review conclusion is to pass but whose review intensity requirement is not lower than the corresponding intensity threshold, and the total amount of review resources occupied. The missed items, falsely blocked items, and total amount of review resources occupied are weighted by the missed penalty coefficient, the falsely blocked penalty coefficient, and the review resource occupied coefficient, respectively, and include a regularization term for the deviation between the scene weight and the scene weight prior. The intensity threshold is solved by one-dimensional search and the scene weight is solved by gradient update iteratively; the interception rate parameter is automatically updated according to sliding window statistics; the calibration also includes updating the uncertainty penalty coefficient and modal weight, where the modal weight is the weight of each audit model in the comprehensive uncertainty; The review comments, reasons for rejection, and publication results are written into the audit chain record for the calculation of the preset calibration loss function. When the time-weighted average of the difference between the quantified value of the manual review conclusion and the corresponding review intensity requirement in the window remains deviated from zero within a preset time period, the time-weighted average is used as an additional term of the preset calibration loss function.

10. An AI-based dynamic review process orchestration system based on multimodal risk scoring, used to execute the AI-based dynamic review process orchestration method as described in any one of claims 1 to 9, characterized in that, include: The feature acquisition module is used to receive the object to be reviewed and its business context, perform multimodal analysis on text, image, video and audio data, and obtain dimensional risk scores for multiple risk dimensions. The risk calculation module is used to perform probability calculations on the category distribution output by the audit model to obtain the comprehensive uncertainty, to sum the risk scores of the dimensions according to the scenario weights to obtain the comprehensive risk score, and to calculate the audit intensity requirement based on the comprehensive risk score and the comprehensive uncertainty adjusted by the uncertainty penalty coefficient. The process orchestration module is used to configure interception rate parameters and resource consumption parameters for candidate review nodes, aggregate and calculate process strength based on the interception rate parameters of each review node, and automatically solve the target review process with the goal of minimizing the total review resource consumption among candidate review processes whose process strength is not less than the review strength requirement and meets the preset hard rules. The process execution module is used to automatically distribute audit tasks to the audit terminals corresponding to each audit node according to the target audit process. During the execution, it continuously acquires related monitoring data and recalculates the audit intensity requirements. When the preset adjustment conditions are met, it determines the remaining intensity requirements based on the process intensity of the executed audit nodes and automatically re-solves the remaining sub-processes. The calibration module is used to collect the conclusions of manual review, update the scene weights, intensity thresholds and interception rate parameters according to the preset calibration loss function, and feed them back to the risk calculation module, process orchestration module and process execution module. The audit log module is used to write the preset adjustment conditions, risk increments, and the audit nodes that are added or canceled and their risk basis into the audit chain record each time the target audit process is adjusted, and to write the audit opinion, rejection reason and release result into the audit chain record.

Images