An artificial intelligence-based content compliance real-time review method and system

Abstract

The application discloses a content compliance real-time review method and system based on artificial intelligence, relates to the technical field of content review, and comprises the following steps: screening out candidate review models that do not meet a preset response time limit, performing model capability bit masking and context demand bit masking inclusion determination on the screened candidate review models, obtaining a target review model, inputting multi-modal content and a dynamic context vector into the target review model to perform joint reasoning, generating a preliminary review determination result, performing result arbitration according to a review category and category confidence in the preliminary review determination result in combination with a dynamic confidence threshold in a dynamic review strategy, and obtaining a final review result; and the application improves the real-time response speed, semantic understanding accuracy and economic efficiency of resource utilization of content compliance review by generating the final review result.

Description

Technical Field

[0001] This invention relates to the field of content review technology, and in particular to a real-time content compliance review method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of internet information technology, especially the popularization of Web 2.0 and social media, user-generated content (UGC) has exploded. Cyberspace has become the main carrier of information dissemination, containing not only massive amounts of text information but also a large amount of multimodal data such as images. In order to maintain a clean and secure online environment and prevent the spread of illegal and harmful information (such as terrorism, vulgarity, and rumors), content compliance review technology has emerged and evolved from manual review to automated machine review. Early content review mainly relied on text filtering technology based on keyword matching and image fingerprinting technology based on hash value comparison. In recent years, with the advancement of deep learning and computer vision technology, single-modal review models based on convolutional neural networks and natural language processing (NLP) have gradually become mainstream, laying a solid technical foundation for building an efficient and accurate online content governance system and significantly improving the efficiency of identifying explicit violations.

[0003] Nevertheless, existing content review methods still have room for improvement. First, they treat the content to be reviewed as isolated samples, ignoring the important influence of the publication context on the determination of content compliance, resulting in a lack of specificity and flexibility in the review results. Second, they usually adopt a strategy of serial use of a single model or parallel use of a fixed model pool, which fails to effectively balance the contradiction between "review accuracy" and "response latency". When facing high-concurrency real-time review scenarios, they cannot dynamically select the most suitable candidate model according to the current urgency level of the handling. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based real-time content compliance review method to address the problem of review results lacking specificity and flexibility.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a real-time content compliance review method based on artificial intelligence, comprising, Collect multimodal content to be reviewed, extract relevant contextual information for feature fusion, and obtain dynamic context vectors; The dynamic context vector is input into the context policy network for policy reasoning to obtain the dynamic review policy, which includes the urgency level of the action, the model version identifier, and the dynamic confidence threshold. Extract candidate review models corresponding to model version identifiers from the pre-deployed model library; Perform requirement mapping processing on the dynamic context vector, and perform bitmask generation operation in combination with the urgency level of the handling to obtain the context requirement bitmask corresponding to the multimodal content to be reviewed; Candidate review models that do not meet the preset response time limit are filtered out, and the inclusion determination of the model capability bitmask and the context requirement bitmask is performed on the filtered candidate review models to obtain the target review model. Multimodal content and dynamic context vectors are input into the target review model to perform joint inference and generate preliminary review judgment results; Based on the review category and category confidence level in the preliminary review results, and combined with the dynamic confidence level threshold in the dynamic review strategy, the results are arbitrated to obtain the final review result.

[0007] As a preferred embodiment of the AI-based real-time content compliance review method of the present invention, the multimodal content to be reviewed includes text content and images.

[0008] As a preferred embodiment of the AI-based real-time content compliance review method of the present invention, the step of obtaining the dynamic context vector specifically includes: Based on the multimodal content to be reviewed, extract the relevant contextual information; Perform category encoding on discrete class information in the context information to obtain category encoding results, and perform numerical normalization on continuous class information in the context information to obtain numerical normalization results; The category encoding result and the numerical normalization result are concatenated to form the initial concatenated feature vector; The initial concatenated feature vectors are linearly mapped to obtain the dynamic context vector.

[0009] As a preferred embodiment of the AI-based real-time content compliance review method of the present invention, the acquisition of the dynamic review strategy specifically includes: The dynamic context vector is input into the context policy network for feature response calculation to obtain the corresponding policy discrimination result; Based on the policy discrimination results, identify the urgency level of the action, the model version identifier, and the dynamic confidence threshold corresponding to the dynamic context vector; By combining the urgency level of the action, the model version identifier, and the dynamic confidence threshold, a dynamic review strategy is obtained.

[0010] As a preferred embodiment of the real-time content compliance review method based on artificial intelligence described in this invention, the step of extracting the candidate review model corresponding to the model version identifier refers to matching the model version identifier with the pre-registered version index information in the model library item by item. When the version index information matches the model version identifier, the review model associated with the corresponding version index information is extracted to obtain the candidate review model.

[0011] As a preferred embodiment of the AI-based real-time content compliance review method of the present invention, the step of obtaining the contextual requirement bitmask corresponding to the multimodal content to be reviewed specifically involves: A linear transformation is performed on the dynamic context vector to obtain complex semantic demand identifier values, context-related demand identifier values, and high propagation risk demand identifier values. Based on the urgency level of the handling, the real-time processing demand identifier value is determined and used as the real-time processing demand bit. The complex semantic demand identifier value, the context-related demand identifier value, and the high propagation risk demand identifier value are compared with the demand threshold to obtain the complex semantic demand bit, the context-related demand bit, and the high propagation risk demand bit; The contextual requirement bit mask is obtained by arranging the bits for complex semantic requirements, contextual association requirements, high propagation risk requirements, and real-time processing requirements.

[0012] As a preferred embodiment of the AI-based real-time content compliance review method of the present invention, the acquisition of the target review model specifically includes: Collect the model inference latency corresponding to each candidate review model, filter out the candidate review models whose model inference latency is greater than the preset response time limit, and obtain the filtered candidate review models. For each candidate review model after screening, perform a bitwise AND operation between the model capability bitmask and the context requirement bitmask. When the bitwise AND operation result is consistent with the context requirement bitmask, the model capability bitmask completely covers the context requirement bitmask. Candidate review models whose model capability bitmasks completely cover the context requirement bitmasks are used as target review models.

[0013] As a preferred embodiment of the AI-based real-time content compliance review method of the present invention, the generation of preliminary review judgment results specifically includes: After inputting the multimodal content and dynamic context vector into the target review model, the target review model performs content feature extraction on the text content and image content in the multimodal content respectively to obtain the corresponding multimodal content features; The target review model performs context feature mapping on dynamic context vectors to obtain context constraint features corresponding to multimodal content features; Multimodal content features and contextual constraint features are jointly correlated and calculated within the target review model to obtain violation discrimination features; Linear mapping and normalization are performed on the violation identification features to obtain the probability values ​​corresponding to each review category; Based on the probability values ​​corresponding to each review category, the review category with the highest probability value is selected as the review category output, and the corresponding probability value is used as the category confidence level. The preliminary review judgment result is obtained by combining the review category output results and the category confidence.

[0014] As a preferred embodiment of the AI-based real-time content compliance review method of the present invention, the step of obtaining the final review result specifically includes: Extract the review category and category confidence from the preliminary review results. When the review category is a violation category, the corresponding category confidence is greater than the dynamic confidence threshold, and the difference between the category confidence and the dynamic confidence threshold is greater than the upper limit of the preset review interval, the multimodal content to be reviewed is determined to be in violation. When the review category is a non-violation category, the corresponding category confidence level is greater than the dynamic confidence level threshold, and the difference between the category confidence level and the dynamic confidence level threshold is greater than the upper limit of the preset review interval, the multimodal content to be reviewed is determined to be compliant. When the difference between the category confidence level and the dynamic confidence level threshold is less than or equal to the upper limit of the preset review interval, the multimodal content to be reviewed is determined to require manual review. The judgment result will be used as the final review result for the multimodal content to be reviewed.

[0015] Secondly, this invention provides an artificial intelligence-based real-time content compliance review system, comprising, The acquisition module collects multimodal content to be reviewed and extracts relevant contextual information for feature fusion to obtain dynamic context vectors. The strategy module inputs dynamic context vectors into the context policy network for policy reasoning to obtain dynamic review policies, which include the urgency level of the action, the model version identifier, and the dynamic confidence threshold. The identification module extracts candidate review models corresponding to model version identifiers from a pre-deployed model library; The calculation module performs requirement mapping processing on the dynamic context vector and performs bitmask generation operation in combination with the urgency level of the handling to obtain the context requirement bitmask corresponding to the multimodal content to be reviewed. The indicator module filters out candidate review models that do not meet the preset response time limit, and performs an inclusion determination of the model capability bitmask and the context requirement bitmask on the filtered candidate review models to obtain the target review model. The reasoning module inputs multimodal content and dynamic context vectors into the target review model to perform joint reasoning and generate preliminary review judgment results; The arbitration module, based on the review category and category confidence level in the preliminary review judgment result, and combined with the dynamic confidence level threshold in the dynamic review strategy, performs result arbitration to obtain the final review result.

[0016] The beneficial effects of this invention are as follows: By constructing a "dynamic context vector" and a "context policy network," the invention achieves automatic generation of review strategies without manual intervention and dynamic routing of models. It also endows the invention with context-based global awareness capabilities and further introduces an inclusion determination mechanism for model capability bitmasks and context requirement bitmasks. Under the premise of meeting real-time constraints, it accurately matches the review model with the lowest complexity and the strongest targeting, effectively solving the imbalance between "accuracy" and "efficiency" in the previous single model architecture. Finally, combined with the result arbitration mechanism of dynamic confidence threshold, it forms a defense closed loop with risk self-adaptation. Without relying solely on computing power, it improves the real-time response speed, semantic understanding accuracy, and resource utilization economy of content compliance review. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart for a real-time content compliance review method based on artificial intelligence.

[0019] Figure 2 A flowchart for obtaining dynamic context vectors.

[0020] Figure 3 A flowchart for obtaining dynamic review strategies.

[0021] Figure 4 A flowchart for obtaining preliminary review and judgment results. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a real-time content compliance review method based on artificial intelligence, including the following steps: S1. Collect multimodal content to be reviewed, extract relevant contextual information for feature fusion, and obtain dynamic context vector.

[0026] S1.1 It should be noted that the system receives publishing data corresponding to multimodal content submitted by user accounts on the content publishing page, which includes title text, body text, and image files corresponding to the image upload area. The system reads the title text and body text from the publishing data as text content and reads the image files as images.

[0027] S1.2 It should be noted that when associating text content and images, the user account identifier, publication serial number, and submission timestamp in the corresponding publication data are read. The combination of the user account identifier, publication serial number, and submission timestamp is used as the content identifier to represent the same publication object. The same publication object refers to the title text, body text, and all image files in the image upload area generated by the same user account in the same publication submission process. For multiple images uploaded simultaneously in the same publication submission process, all images are associated with the corresponding title text and body text to the same content identifier. When the multimodal content to be reviewed is a new submission formed by editing and republishing the original publication, adding supplementary images, restoring a draft and resubmitting, or forwarding and referencing it, the combination of the user account identifier, publication serial number, and submission timestamp corresponding to the new submission is used as the basis for generating a new content identifier, and the corresponding content is determined as the new publication object. Based on the content identifier, the corresponding publishing records and publisher account records are located. The publishing time and initial publishing location are extracted from the publishing records, and the publisher identity information and historical publishing behavior information are extracted from the publisher account records. When the multimodal content to be reviewed has been published and formed interaction records and dissemination records, the interaction change records and dissemination change records after publication are read. The interaction change records include the number of views, likes, comments, reposts, and reports at different times. The dissemination change records include the dissemination scope, dissemination level, and dissemination duration at different times. The publishing time, initial publishing location, publisher identity information, historical publishing behavior information, and the read interaction change records and dissemination change records are arranged and organized in chronological order to obtain the context information associated with the multimodal content to be reviewed.

[0028] S1.3. It should be noted that the time period to which the publication belongs, the category of the initial publication location, the category of the publisher's identity, the category of historical publication behavior, and the category of dissemination level in the context information are classified and read, and a corresponding category number is assigned to each category. Each category number is used as the category coding result corresponding to the discrete category information. The number of views, likes, comments, reposts, reports, dissemination scope, dissemination duration, and number of historical violations in the context information are numerically read, and the maximum and minimum values ​​of each continuous category information in the current processing batch are calculated. Min-max normalization is performed on each continuous category information to obtain the normalized value corresponding to each continuous category information. The category coding result and the normalized value are combined sequentially according to the correspondence of the same content identifier to obtain the initial spliced ​​feature vector (13 dimensions). The initial spliced ​​feature vector is multiplied by the mapping weight matrix (dimension 8×13), and a bias vector (8 dimensions) is superimposed to obtain the dynamic context vector.

[0029] Furthermore, the mapping weight matrix and bias vector are set to transform the initial concatenated feature vector into a dynamic context vector suitable for subsequent contextual policy network processing. In one embodiment, the element values ​​of the mapping weight matrix can be set to a range of -1.0 to 1.0, and the element values ​​of the bias vector can be set to a range of -0.5 to 0.5. This can reduce the dimensionality redundancy of the initial concatenated feature vector while retaining the main features of publication time, publication location, publisher status, interaction change status, and propagation change status, thereby improving the processing efficiency of the subsequent contextual policy network and enhancing the ability of the dynamic context vector to represent the current review context.

[0030] S2. Input the dynamic context vector into the context policy network for policy reasoning to obtain the dynamic review policy.

[0031] S2.1 It should be noted that historical review samples are collected and historical dynamic context vectors are constructed; the actual urgency level of the handling, the actual model version identifier, and the actual dynamic confidence threshold used in the historical review records are used as supervision labels; the historical dynamic context vectors and supervision labels together constitute the training set.

[0032] The Contextual Policy Network (CPR) is a multi-layer feedforward neural network architecture, comprising an input layer, two hidden computation layers, and a three-branch policy output layer. The input layer output dimension is 32, the first hidden computation layer output dimension is 16, the second hidden computation layer output dimension is 16, the urgency level output dimension is 3, the model version identifier output dimension is 3, and the dynamic confidence threshold output dimension is 1. Correspondingly, the input layer weight matrix has a dimension of 32×8, the input layer bias vector has a dimension of 32, and the first hidden computation layer weight matrix has a dimension of 16×32. The dimensions of the hidden computation layer bias vector are set to 16, the second hidden computation layer weight matrix is ​​set to 16×16, the second hidden computation layer bias vector is set to 16, the urgency level output branch weight matrix is ​​set to 3×16, the urgency level output branch bias vector is set to 3, the model version identifier output branch weight matrix is ​​set to 3×16, the model version identifier output branch bias vector is set to 3, the dynamic confidence threshold output branch weight matrix is ​​set to 1×16, and the dynamic confidence threshold output branch bias vector is set to 1. These layer dimensions and corresponding parameters are set to ensure that, while retaining the main contextual features of the dynamic contextual vector, feature expansion, feature compression, feature extraction, and multi-task output are performed layer by layer. This allows the contextual policy network to simultaneously output policy discrimination results corresponding to the urgency level, model version identifier, and dynamic confidence threshold.

[0033] The training set is input into the context policy network for forward propagation to obtain the predicted policy. The predicted policy is compared with the supervision label to calculate the loss value. The loss value is used as the basis for parameter optimization. The gradient of each learnable parameter in the context policy network is calculated using the backpropagation algorithm. Then, the learnable parameters are updated according to the gradient descent method. The forward propagation, loss value calculation and parameter update process is repeated until the preset maximum number of iterations is reached to obtain the trained context policy network.

[0034] The maximum number of iterations is set to impose an upper limit on the number of parameter update rounds of the context policy network. In one embodiment, the maximum number of iterations can be set to 30 times to avoid the context policy network from learning the correspondence between dynamic context vectors and supervision labels insufficiently due to too few iterations, while avoiding the training time increase and the risk of overfitting increase due to too many iterations.

[0035] S2.2 It should be noted that the dynamic context vector is input into the trained context policy network. The input layer multiplies the dynamic context vector with the input layer weight matrix, sums the product results, and then superimposes the input layer bias vector to obtain the linear computation result of the input layer. The ReLU function is applied to the linear computation result of the input layer to obtain the input layer feature representation. This feature representation is then input into the first hidden computation layer, where it is multiplied with the first hidden computation layer weight matrix, and the corresponding bias vector is superimposed to obtain the linear output result of the first hidden computation layer. The ReLU function is applied to the linear output result of the first hidden computation layer to obtain the feature representation of the first hidden computation layer. Finally, the feature representation of the first hidden computation layer is input into the second hidden computation layer, where linear transformation and nonlinear activation processing are performed to obtain the shared feature representation.

[0036] The shared feature representations are input into the urgency level output branch, the model version identifier output branch, and the dynamic confidence threshold output branch, respectively. The urgency level output branch performs matrix multiplication on the shared feature representations and the corresponding weight matrices, and then adds a bias vector to obtain a linear output result for the urgency level. The Softmax function is applied to this linear output result to obtain the probability values ​​(including high, medium, and low urgency levels) for each urgency level. Similarly, the model version identifier output branch performs a linear transformation on the shared feature representations and applies the Softmax function to obtain the probability values ​​corresponding to each model version identifier. Finally, the dynamic confidence threshold output branch performs a linear transformation on the shared feature representations and applies the Sigmoid function to obtain a dynamic confidence threshold between zero and one.

[0037] Based on the high-level, medium-level, and low-level probability values ​​output by the urgency level output branch, the level with the highest probability value is selected as the urgency level output result. Based on the probability values ​​corresponding to each model version identifier, the model version identifier output result with the highest probability value is selected. The urgency level output result, the model version identifier output result, and the dynamic confidence threshold are combined to obtain the dynamic review strategy corresponding to the dynamic context vector. The dynamic review strategy includes the urgency level, model version identifier, and dynamic confidence threshold.

[0038] S3. Extract candidate review models corresponding to model version identifiers from the pre-deployed model library.

[0039] It should be noted that the process involves reading the model version identifier from the dynamic review strategy and the version index information stored for each review model in the model library; performing a step-by-step comparison between the model version identifier and each version index information according to the registration order of the version index information to determine whether the model version identifier is consistent with the version index field in each version index information; when the model version identifier is inconsistent with the currently compared version index information, the process continues to read the next version index information and perform subsequent step-by-step comparisons; when the model version identifier is consistent with a certain version index information, the review model record that establishes a relationship with the corresponding version index information is located, and the review model identifier, review model availability status, and review model capability description are extracted from the review model record; the review model availability status is read, and when the review model availability status indicates that the review model is in a callable state, the corresponding review model is retained to obtain a candidate review model; when all version index information has been compared and there is no version index information consistent with the model version identifier, the default review model record pre-registered in the model library is read, and when the default review model is in a callable state, the default review model is determined as a candidate review model.

[0040] The model library is constructed as follows: historical review samples are collected, including text content, images, and corresponding violation category annotations, handling result annotations, and review scenario annotations; multiple review models are constructed or deployed based on the historical review samples, such as a lightweight image and text review model, an image and text compliance review model, and a high-precision semantic association review model; the multiple review models differ in network scale, inference latency, supported violation category range, applicable content complexity level, or deployment location; when conducting offline verification tests on each review model, a complex semantic test subset, a context association test subset, and a high propagation risk test subset are first constructed based on the historical review samples; among them, the complex semantic test subset includes samples corresponding to the implicit semantics of long texts, the semantics of avoidance expressions, and the semantics of cross-sentence combinations; the context association test subset includes samples containing publishing records, publisher historical behavior information, and propagation change information; and the high propagation risk test subset includes samples corresponding to high-popularity propagation, high-level propagation, and high-report growth scenarios. For complex semantic review capabilities, the macro-average F1 score and recall rate of the corresponding review model on the complex semantic test subset are used as evaluation indicators. When the macro-average F1 score is greater than or equal to the first capability threshold and the recall rate is greater than or equal to the second capability threshold, the capability position corresponding to the complex semantic review capability is set to 1; otherwise, it is set to 0.

[0041] For context-related review capabilities, text content, images and posting records, publisher historical behavior information, and dissemination change information are respectively input into the content feature extraction branch and the context feature extraction branch, and the output results of the two branches are fused and judged. When the macro average F1 value of the corresponding review model on the context-related test subset is greater than or equal to the third capability threshold, the capability position corresponding to the context-related review capability is set to 1, otherwise it is set to 0.

[0042] For response capabilities in high-risk scenarios, the macro-average F1 score and high-risk category recall rate of the corresponding review model on the high-risk test subset are used as evaluation indicators. When the macro-average F1 score is greater than or equal to the fourth capability threshold and the high-risk category recall rate is greater than or equal to the fifth capability threshold, the capability position corresponding to the high-risk scenario response capability is set to 1; otherwise, it is set to 0. For real-time inference capability, under the conditions of a preset hardware environment (e.g., NVIDIA T4), a fixed batch size of 1, and a fixed range of input resolution and text length, the average inference latency of the corresponding review model performing one inference on a single sample to be reviewed is calculated. When the average inference latency is less than or equal to the fast latency threshold, the capability position corresponding to the real-time inference capability is set to 1; otherwise, it is set to 0.

[0043] The capability bits corresponding to complex semantic review capabilities, contextual review capabilities, high-propagation-risk scenario response capabilities, and real-time inference capabilities are arranged to obtain the model capability bit mask for each review model. Based on the resource consumption and structural complexity of each review model, a model level is pre-registered for each review model, and the corresponding model inference latency is recorded. After establishing a correlation between the model capability bit mask, model level, and model inference latency and each review model record, a fallback review model is selected from multiple review models as the default review model. The default review model is used to perform fallback extraction of candidate review models when the model version identifier does not match the corresponding version index information. The review model identifier, default model tag, review model availability status, and review model deployment information corresponding to the default review model are associated and registered to form a default review model record, which is stored in the model library along with the remaining review models.

[0044] Furthermore, a review model being in a callable state means that the version index information, model capability bitmask, model level, and model inference latency of the corresponding review model have been registered and associated with the review model record, the corresponding review model has been deployed on an available inference node, the inference service process is running normally, the corresponding computing resources meet the minimum inference requirements, and the corresponding review model has not been disabled, has not experienced loading failures, or service anomalies. When all the above conditions are met, the review model availability status of the corresponding review model is determined to be callable; otherwise, it is determined to be uncallable.

[0045] The first capability threshold is set to determine whether the corresponding review model has the comprehensive recognition capability to meet the expected requirements on the complex semantic test subset. In one embodiment, the first capability threshold can be set to 0.85 to avoid recording a review model with insufficient recognition capability for implicit semantics of long texts, semantic avoidance, and semantic cross-sentence combination as having complex semantic review capability.

[0046] The second capability threshold is set to determine whether the corresponding review model has sufficient detection capability for the target violation sample on a subset of complex semantic tests; in one embodiment, the second capability threshold can be set to 0.80.

[0047] The third capability threshold is set to determine whether the corresponding review model has the contextual recognition capability to meet the expected requirements when performing joint judgment by combining the posting record, the publisher's historical behavior information and the dissemination change information; in one embodiment, the third capability threshold can be set to 0.82.

[0048] The fourth capability threshold is set to determine whether the corresponding review model has the comprehensive violation identification capability to meet the expected requirements on the high-propagation-risk test subset; in one embodiment, the fourth capability threshold can be set to 0.83.

[0049] The fifth capability threshold is set to determine whether the corresponding review model has sufficient detection capability for high-risk categories on the high-transmission risk test subset; in one embodiment, the fifth capability threshold can be set to 0.85.

[0050] The fast latency threshold is set to determine whether the corresponding review model has real-time inference capability in the current deployment environment. In one embodiment, the fast latency threshold can be set to 80 milliseconds. 80 milliseconds is the upper limit of the average inference latency measured when performing one inference on a single sample to be reviewed under the conditions of GPU model NVIDIA T4, batch size of 1, text length not exceeding 512 words and image input resolution of 224×224. It is used to avoid classifying review models with slow inference response as having real-time inference capability.

[0051] The lightweight image and text review model has a simple network structure, small parameter size, and short inference latency; therefore, the corresponding model level is set to 1. The image and text compliance review model also includes text. The Transformer encoding layer, image vision Transformer encoding layer, context mapping layer, and cross-modal fusion layer have higher resource consumption and structural complexity than the lightweight image and text review model, thus corresponding to the intermediate model level, such as 2; the high-precision semantic association review model further performs more complex association determination on the basis of image and text alignment calculation and cross-modal deep fusion, with higher resource consumption and more complex structure, thus corresponding to the high model level, such as 3.

[0052] The lightweight image and text review model includes a shallow text encoding layer, a lightweight image convolutional encoding layer, a feature concatenation layer, and a binary classification output layer to output violation / non-violation binary classification results. The image and text compliance review model includes a text Transformer encoding layer, an image visual Transformer encoding layer, a context mapping layer, a cross-modal fusion layer, and a multi-classification output layer to output the probability distributions corresponding to multiple violation categories. The high-precision semantic association review model includes a text Transformer encoding layer, an image visual Transformer encoding layer, an image-text alignment calculation layer, a cross-modal deep fusion layer, and dual output branches to output violation category results and image-text association risk results. It should be noted that the lightweight image and text review model, the general image and text compliance review model, and the high-precision semantic association review model are only exemplary examples of review models in the model library. Other review models can be configured in the model library according to actual review needs to adapt to review tasks with different violation category ranges, different content complexity levels, different inference latency requirements, or different deployment locations.

[0053] S4. Perform requirement mapping processing on the dynamic context vector, and perform bitmask generation operation in combination with the urgency level of the handling to obtain the context requirement bitmask corresponding to the multimodal content to be reviewed.

[0054] S4.1 It should be noted that the pre-configured demand mapping weight matrix and demand mapping bias vector are read. The demand mapping weight matrix has a dimension of 3×8, and the demand mapping bias vector has a dimension of 3. The demand mapping weight matrix and the dynamic context vector are multiplied together, and the demand mapping bias vector is then superimposed to obtain a 3-dimensional demand response vector. The first dimension of the demand response vector is determined as the complex semantic demand identifier value, the second dimension as the context-related demand identifier value, and the third dimension as the high propagation risk demand identifier value. The real-time processing demand identifier value is determined according to the urgency level in the dynamic review strategy. When the urgency level is high, the real-time processing demand identifier value is set to 1; when the urgency level is medium or low, the real-time processing demand identifier value is set to 0. The complex semantic requirement identifier, context-related requirement identifier, and high propagation risk requirement identifier are compared with the requirement threshold. When the corresponding requirement identifier is greater than or equal to the requirement threshold, the corresponding requirement bit is set to 1; otherwise, it is set to 0. The real-time processing requirement identifier is determined as the real-time processing requirement bit. The complex semantic requirement bit, context-related requirement bit, high propagation risk requirement bit, and real-time processing requirement bit are arranged to obtain the context requirement bit mask.

[0055] Furthermore, the demand mapping weight matrix and demand mapping bias vector are set to map the various dimensions of the dynamic context vector into complex semantic demand identifier values, context-related demand identifier values, and high-propagation-risk demand identifier values.

[0056] The complex semantic requirement bit, contextual requirement bit, high propagation risk requirement bit, and real-time processing requirement bit correspond to the first, second, third, and fourth bits in the context requirement bit mask, respectively; each capability bit in the model capability bit mask corresponds one-to-one with each requirement bit in the context requirement bit mask using the same bit order.

[0057] The demand threshold is set to identify whether the multimodal content to be reviewed currently poses a clear demand for the corresponding review capabilities. In one embodiment, the demand threshold can be set to 0.5, which can avoid mislabeling weak demand situations as strong demand situations while retaining contextual demand sensitivity, thereby improving the stability of the subsequent target review model selection process.

[0058] S5. Filter out candidate review models that do not meet the preset response time limit, and perform inclusion determination of the model capability bitmask and the context requirement bitmask on the filtered candidate review models to obtain the target review model.

[0059] S5.1 It should be noted that the corresponding preset response time limit is determined according to the urgency level of the handling. For example, based on the urgency level, a pre-established correspondence between the urgency level and the response time limit is established. Specifically, when the urgency level is high, the corresponding preset response time limit can be set to 500 milliseconds; when the urgency level is medium, the corresponding preset response time limit can be set to 1500 milliseconds; and when the urgency level is low, the corresponding preset response time limit can be set to 3000 milliseconds. Using the above settings allows high-risk, high-urgency multimodal content to be reviewed to be preferentially matched with candidate review models with shorter inference latency, while reserving more inference time for medium-urgency and low-urgency multimodal content to be reviewed, thus balancing real-time review efficiency and review accuracy. The model inference latency corresponding to each candidate review model is compared with the preset response time limit one by one. When the model inference latency corresponding to a candidate review model is less than or equal to the preset response time limit, the corresponding candidate review model is retained; when the model inference latency corresponding to a candidate review model is greater than the preset response time limit, the corresponding candidate review model is eliminated, resulting in the filtered candidate review models.

[0060] S5.2 It should be noted that for each candidate review model after screening, the corresponding model capability bitmask is read, and a bitwise AND operation is performed between the model capability bitmask and the context requirement bitmask. When the bitwise AND result is completely consistent with the context requirement bitmask, it is determined that the model capability bitmask of the corresponding candidate review model completely covers the context requirement bitmask. If one or more candidate review models among the screened candidate review models have model capability bitmasks that completely cover the context requirement bitmask, the model level of the corresponding candidate review model is read, and the candidate review model with the lowest model level is determined as the target review model. When there are two or more candidate review models with the same model level, the candidate review model with the shortest model inference latency is determined as the target review model.

[0061] Furthermore, if none of the candidate review models selected have a model capability bitmask that completely covers the context requirement bitmask, then the number of bits with a value of 1 in the context requirement bitmask and a value of 0 in the corresponding position in the model capability bitmask is counted for each candidate review model to obtain the coverage gap value corresponding to each candidate review model; the candidate review model with the smallest coverage gap value is determined as the target review model; when there are two or more candidate review models with the same coverage gap value, the candidate review model with the shortest model inference latency is determined as the target review model.

[0062] The preset response time limit is set to ensure that the target review model meets the real-time processing requirements of the current review scenario.

[0063] S6. Input the multimodal content and dynamic context vector into the target review model to perform joint reasoning and generate preliminary review judgment results.

[0064] S6.1 It should be noted that, in one embodiment, when the target review model is a text and image compliance review model, the text Transformer encoding layer of the text and image compliance review model includes 4 encoding sub-layers, and each encoding sub-layer has 4 attention heads; the image and vision Transformer encoding layer includes 4 encoding sub-layers, and each encoding sub-layer has 4 attention heads; the context mapping layer adopts a two-layer fully connected mapping structure, wherein the first layer output dimension is set to 64 dimensions and the second layer output dimension is set to 256 dimensions; the cross-modal fusion layer includes two fusion mapping layers; the multi-classification output layer output dimension is set to 6 dimensions, corresponding to 6 review categories.

[0065] The training process for the text and image compliance review model is as follows: Historical review samples are collected, including text content, image content, historical dynamic context vectors, and corresponding review category labels corresponding to historically published data. The review category labels are determined by historical review conclusions, and the review categories include vulgar and illegal content, terrorism and violence, rumors, hate speech, advertising and traffic redirection, and non-illegal content. The text content, image content, and historical dynamic context vectors from the historical review samples are used as training samples, and the review category labels are used as supervision labels, together forming the training set for the text and image compliance review model.

[0066] S6.2 It should be noted that the training set is input into the image and text compliance review model to perform forward propagation calculation, obtain the prediction results, compare the prediction results with the supervision labels, and calculate the classification loss value. The classification loss value is used as the basis for parameter optimization. The gradient of each learnable parameter in the image and text compliance review model is calculated using the backpropagation algorithm, and the gradient descent method is used to update each learnable parameter. The forward propagation, loss value calculation, and parameter update process is repeated until the preset maximum number of iterations is reached, resulting in the trained image and text compliance review model. Setting the maximum number of iterations to 50 rounds allows the image and text compliance review model to ensure sufficient training while controlling training time and the risk of overfitting, thereby improving the stability and accuracy of the trained image and text compliance review model in determining the violation category.

[0067] S6.3. It should be noted that the multimodal content and dynamic context vectors are input into the trained text-image compliance review model. The text Transformer encoding layer performs word segmentation on the text content and organizes the text content into a text sequence with a length not exceeding 128 words. The text Transformer encoding layer maps each word to a 256-dimensional word vector and superimposes the corresponding 256-dimensional positional encoding to obtain the text input sequence. The text input sequence is then sequentially input into four text Transformer encoding sub-layers. Within each text Transformer encoding sub-layer, a linear transformation is performed on the current input sequence using the query mapping weight matrix, key mapping weight matrix, and value mapping weight matrix, respectively, to obtain the query vector sequence, key vector sequence, and value vector sequence. A dot product operation is performed on the query vector sequence and the key vector sequence to obtain the association score between each word. The association score is then... The attention weights are obtained by applying the Softmax function. The attention weights are then used to perform a weighted summation on the value vector sequence to obtain the attention output. The attention output is then summed with the residual of the current input sequence and subjected to layer normalization. The normalized result is input into the feedforward layer, where a linear transformation is performed on the first feedforward layer weight matrix, and a bias vector is superimposed. The ReLU function is then applied to obtain the first feedforward layer output. The first feedforward layer output is then summed with the second feedforward layer weight matrix, and a bias vector is superimposed to obtain the feedforward output. The feedforward output is then summed with the feedforward layer input and subjected to layer normalization again to obtain the output of the current text Transformer encoding sublayer. After all four text Transformer encoding sublayers have been executed, average pooling is performed along the sequence dimension on the output text semantic sequence features to obtain 256-dimensional text features.

[0068] S6.4 It should be noted that the image visual Transformer coding layer adjusts the image content to 224×224 pixels, and then divides it into 196 image blocks of 16×16 pixels; each image block is flattened into a 768-dimensional pixel vector, and a linear transformation is performed using a 256×768 image block mapping weight matrix and the corresponding bias vector to obtain a 256-dimensional image block vector corresponding to each image block; the image visual Transformer coding layer reads the positional encoding corresponding to each image block from the 196×256 image positional encoding matrix, and adds it element-wise with each image block vector to obtain the image input sequence; the image input sequence is sequentially input into four image visual Transformer coding sub-layers, and in each image visual Transformer coding sub-layer, the same self-attention coding processing as the text content is performed to obtain 256-dimensional image features.

[0069] S6.5 It should be noted that the first context mapping layer is calculated on the dynamic context vector. The weight matrix of the first context mapping layer is multiplied with the dynamic context vector, and a bias vector is superimposed to obtain a 64-dimensional intermediate context representation. The ReLU function is applied to the 64-dimensional intermediate context representation to obtain the output result of the first context mapping layer. The weight matrix of the second context mapping layer is multiplied with the output result of the first context mapping layer, and a bias vector is superimposed to obtain a 256-dimensional context constraint feature. After applying the ReLU function to the 256-dimensional context constraint feature, it is concatenated with the 256-dimensional text feature and the 256-dimensional image feature to form a 768-dimensional fused input vector.

[0070] S6.6. It should be noted that the first fusion mapping layer weight matrix is ​​multiplied by the 768-dimensional fusion input vector, and a bias vector is superimposed to obtain a 512-dimensional intermediate fusion representation; the ReLU function is applied to the 512-dimensional intermediate fusion representation to obtain the output result of the first fusion mapping layer; the second fusion mapping layer weight matrix is ​​multiplied by the first fusion mapping layer output result, and a bias vector is superimposed to obtain a 256-dimensional violation discrimination feature; the ReLU function is applied to the 256-dimensional violation discrimination feature, and then multiplied by the multi-class output layer weight matrix, and a bias vector is superimposed to obtain a 6-dimensional linear output result; the 6-dimensional linear output result should be... The Softmax function is used to obtain the probability values ​​corresponding to the categories of vulgarity and violation, terrorism and violence, rumors, hate speech, advertising and traffic redirection, and non-violation. The category with the highest probability value is selected as the category output result, and the corresponding probability value is determined as the category confidence score. When the category output result belongs to the categories of vulgarity and violation, terrorism and violence, rumors, hate speech, or advertising and traffic redirection, the category output result is recorded as a violation category, and the violation category and the category confidence score are combined as the preliminary review judgment result. When the category output result is a non-violation category, the non-violation category and the corresponding category confidence score are combined as the preliminary review judgment result.

[0071] It should be noted that S6.1 uses the target review model as an example of the graphic compliance review model to explain the structural composition, training method and output format of the target review model. The purpose is to illustrate the implementation process of the target review model performing joint inference and generating preliminary review judgment results in step S6.

[0072] S7. Based on the review category and category confidence level in the preliminary review judgment result, and combined with the dynamic confidence level threshold in the dynamic review strategy, perform result arbitration to obtain the final review result.

[0073] It should be noted that, based on the review category and category confidence level in the preliminary review results, when the review category is a violation category, the corresponding category confidence level is greater than the dynamic confidence level threshold, and the difference between the category confidence level and the dynamic confidence level threshold is greater than the upper limit of the preset review interval, the multimodal content to be reviewed is determined to be violation. When the review category is a non-violation category, the corresponding category confidence level is greater than the dynamic confidence level threshold, and the difference between the category confidence level and the dynamic confidence level threshold is greater than the upper limit of the preset review interval, the multimodal content to be reviewed is determined to be compliant. When the difference between the category confidence level and the dynamic confidence level threshold is less than or equal to the upper limit of the preset review interval, the multimodal content to be reviewed is determined to require manual review (if the category confidence level is less than or equal to the dynamic confidence level threshold, the multimodal content to be reviewed is also determined to require manual review). The judgment result (violation, compliance, or requiring manual review) is taken as the final review result of the multimodal content to be reviewed.

[0074] Furthermore, the dynamic confidence threshold is the output value obtained by performing a linear transformation on the shared feature representation and processing it with the Sigmoid function by the dynamic confidence threshold output branch of the contextual policy network in S2.2.

[0075] The upper limit of the review interval is set to determine whether the difference between the category confidence level and the dynamic confidence level threshold meets the judgment condition. In one embodiment, the upper limit of the review interval can be set to 0.1, which can ensure the security of the review of boundary samples while avoiding too many highly certain samples from entering the manual review process, thereby balancing the stability of the review results and the efficiency of real-time processing.

[0076] This embodiment also provides an artificial intelligence-based real-time content compliance review system, including: The acquisition module collects multimodal content to be reviewed and extracts relevant contextual information for feature fusion to obtain dynamic context vectors. The strategy module inputs dynamic context vectors into the context policy network for policy reasoning to obtain dynamic review policies, which include the urgency level of the action, the model version identifier, and the dynamic confidence threshold. The identification module extracts candidate review models corresponding to model version identifiers from a pre-deployed model library; The calculation module performs requirement mapping processing on the dynamic context vector and performs bitmask generation operation in combination with the urgency level of the handling to obtain the context requirement bitmask corresponding to the multimodal content to be reviewed. The indicator module filters out candidate review models that do not meet the preset response time limit, and performs an inclusion determination of the model capability bitmask and the context requirement bitmask on the filtered candidate review models to obtain the target review model. The reasoning module inputs multimodal content and dynamic context vectors into the target review model to perform joint reasoning and generate preliminary review judgment results; The arbitration module, based on the review category and category confidence level in the preliminary review judgment result, and combined with the dynamic confidence level threshold in the dynamic review strategy, performs result arbitration to obtain the final review result.

[0077] This embodiment also provides a computer device applicable to the real-time content compliance review method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the real-time content compliance review method based on artificial intelligence as proposed in the above embodiment.

[0078] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0079] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the real-time content compliance review method based on artificial intelligence as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0080] In summary, this invention achieves automatic generation of review strategies and dynamic model routing without manual intervention by constructing a "dynamic context vector" and a "context policy network." It also endows the system with context-based global awareness capabilities and further introduces an inclusion determination mechanism for model capability bitmasks and context requirement bitmasks. Under the premise of meeting real-time constraints, it accurately matches the review model with the lowest complexity and highest targeting, effectively resolving the imbalance between "accuracy" and "efficiency" in previous single-model architectures. Finally, combined with a result arbitration mechanism based on dynamic confidence thresholds, it forms a risk-adaptive defense closed loop. This improves the real-time response speed, semantic understanding accuracy, and resource utilization economy of content compliance review without relying solely on computational power.

[0081] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time content compliance review method based on artificial intelligence, characterized in that: include, Collect multimodal content to be reviewed, extract relevant contextual information for feature fusion, and obtain dynamic context vectors; The dynamic context vector is input into the context policy network for policy reasoning to obtain the dynamic review policy, which includes the urgency level of the action, the model version identifier, and the dynamic confidence threshold. Extract candidate review models corresponding to model version identifiers from the pre-deployed model library; Perform requirement mapping processing on the dynamic context vector, and perform bitmask generation operation in combination with the urgency level of the handling to obtain the context requirement bitmask corresponding to the multimodal content to be reviewed; Candidate review models that do not meet the preset response time limit are filtered out, and the inclusion determination of the model capability bitmask and the context requirement bitmask is performed on the filtered candidate review models to obtain the target review model; Multimodal content and dynamic context vectors are input into the target review model to perform joint inference and generate preliminary review judgment results; Based on the review category and category confidence level in the preliminary review results, and combined with the dynamic confidence level threshold in the dynamic review strategy, the results are arbitrated to obtain the final review result.

2. The real-time content compliance review method based on artificial intelligence as described in claim 1, characterized in that: The multimodal content to be reviewed includes text content and images.

3. The real-time content compliance review method based on artificial intelligence as described in claim 2, characterized in that: The acquisition of the dynamic context vector specifically involves: Based on the multimodal content to be reviewed, extract the relevant contextual information; Perform category encoding on discrete class information in the context information to obtain category encoding results, and perform numerical normalization on continuous class information in the context information to obtain numerical normalization results; The category encoding result and the numerical normalization result are concatenated to form the initial concatenated feature vector; The initial concatenated feature vectors are linearly mapped to obtain the dynamic context vector.

4. The real-time content compliance review method based on artificial intelligence as described in claim 3, characterized in that: The specific method for obtaining the dynamic review strategy is as follows: The dynamic context vector is input into the context policy network for feature response calculation to obtain the corresponding policy discrimination result; Based on the policy discrimination results, identify the urgency level of the action, the model version identifier, and the dynamic confidence threshold corresponding to the dynamic context vector; By combining the urgency level of the action, the model version identifier, and the dynamic confidence threshold, a dynamic review strategy is obtained.

5. The real-time content compliance review method based on artificial intelligence as described in claim 4, characterized in that: The extraction of candidate review models corresponding to model version identifiers refers to matching the model version identifier with the pre-registered version index information in the model library item by item. When the version index information matches the model version identifier, the review model associated with the corresponding version index information is extracted as a candidate review model.

6. The real-time content compliance review method based on artificial intelligence as described in claim 5, characterized in that: The specific steps for obtaining the contextual requirement bitmask corresponding to the multimodal content to be reviewed are as follows: A linear transformation is performed on the dynamic context vector to obtain complex semantic demand identifier values, context-related demand identifier values, and high propagation risk demand identifier values. Based on the urgency level of the handling, the real-time processing demand identifier value is determined and used as the real-time processing demand bit. The complex semantic demand identifier value, the context-related demand identifier value, and the high propagation risk demand identifier value are compared with the demand threshold to obtain the complex semantic demand bit, the context-related demand bit, and the high propagation risk demand bit; Arrange the bits for complex semantic requirements, contextual requirements, high propagation risk requirements, and real-time processing requirements to obtain the contextual requirement bit mask.

7. The real-time content compliance review method based on artificial intelligence as described in claim 6, characterized in that: The target review model is specifically as follows: Collect the model inference latency corresponding to each candidate review model, filter out the candidate review models whose model inference latency is greater than the preset response time limit, and obtain the filtered candidate review models. For each candidate review model after screening, perform a bitwise AND operation between the model capability bitmask and the context requirement bitmask. When the bitwise AND operation result is consistent with the context requirement bitmask, the model capability bitmask completely covers the context requirement bitmask. Candidate review models whose model capability bitmasks completely cover the context requirement bitmasks are used as target review models.

8. The real-time content compliance review method based on artificial intelligence as described in claim 7, characterized in that: The generation of the preliminary review and judgment result is specifically as follows: After inputting the multimodal content and dynamic context vector into the target review model, the target review model performs content feature extraction on the text content and image content in the multimodal content respectively to obtain the corresponding multimodal content features; The target review model performs context feature mapping on dynamic context vectors to obtain context constraint features corresponding to multimodal content features; Multimodal content features and contextual constraint features are jointly correlated and calculated within the target review model to obtain violation discrimination features; Linear mapping and normalization are performed on the violation identification features to obtain the probability values ​​corresponding to each review category; Based on the probability values ​​corresponding to each review category, the review category with the highest probability value is taken as the review category output result, and the corresponding probability value is taken as the category confidence level; The preliminary review judgment result is obtained by combining the review category output results and the category confidence.

9. The real-time content compliance review method based on artificial intelligence as described in claim 8, characterized in that: The specific steps for obtaining the final review result are as follows: Extract the review category and category confidence from the preliminary review results. When the review category is a violation category, the corresponding category confidence is greater than the dynamic confidence threshold, and the difference between the category confidence and the dynamic confidence threshold is greater than the upper limit of the preset review interval, the multimodal content to be reviewed is determined to be in violation. When the review category is a non-violation category, the corresponding category confidence level is greater than the dynamic confidence level threshold, and the difference between the category confidence level and the dynamic confidence level threshold is greater than the upper limit of the preset review interval, the multimodal content to be reviewed is determined to be compliant. When the difference between the category confidence level and the dynamic confidence level threshold is less than or equal to the upper limit of the preset review interval, the multimodal content to be reviewed is determined to require manual review. The judgment result will be used as the final review result for the multimodal content to be reviewed.

10. A real-time content compliance review system based on artificial intelligence, based on the real-time content compliance review method based on artificial intelligence as described in any one of claims 1 to 9, characterized in that: include, The acquisition module collects multimodal content to be reviewed and extracts relevant contextual information for feature fusion to obtain dynamic context vectors. The strategy module inputs dynamic context vectors into the context policy network for policy reasoning to obtain dynamic review policies, which include the urgency level of the action, the model version identifier, and the dynamic confidence threshold. The identification module extracts candidate review models corresponding to model version identifiers from a pre-deployed model library; The calculation module performs requirement mapping processing on the dynamic context vector and performs bitmask generation operation in combination with the urgency level of the handling to obtain the context requirement bitmask corresponding to the multimodal content to be reviewed. The indicator module filters out candidate review models that do not meet the preset response time limit, and performs an inclusion determination of the model capability bitmask and the context requirement bitmask on the filtered candidate review models to obtain the target review model. The reasoning module inputs multimodal content and dynamic context vectors into the target review model to perform joint reasoning and generate preliminary review judgment results; The arbitration module, based on the review category and category confidence level in the preliminary review judgment result, and combined with the dynamic confidence level threshold in the dynamic review strategy, performs result arbitration to obtain the final review result.

Images