User generated content auditing method and device based on large model, equipment and medium

Abstract

The invention discloses a user generated content auditing method and device based on a large model, equipment and a medium, and relates to the technical field of information security, and the method comprises the steps: obtaining multi-modal data of a target instance on a target platform through a terminal SDK, processing the multi-modal data through an edge node, and transmitting the obtained processed data to a central cloud; in the central cloud, determining a triple based on the target features of the processed data through a large model, and determining the processed data of which the violation confidence is in a preset confidence interval as a target sample; in the central cloud, converting the triple into a resource description framework diagram, and if the resource description framework diagram is matched with a rule in a preset rule knowledge base, generating a corresponding violation logic; and in the central cloud, checking the target sample based on the violation logic through a preset multi-modal model. Therefore, the efficiency, accuracy and flexibility of auditing the content generated by the user can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information security, and in particular to a user generated content auditing method and device based on a large model, equipment and medium. BACKGROUND

[0002] In the context of the growth of digital content, the scale of UGC (User Generated Content) is expanding exponentially, and the content form of multi-modal and cross-platform continues to evolve. The platform needs to process hundreds of billions of pieces of data such as text, pictures, and audio and video every day in real time, which puts high requirements on content auditing technology.

[0003] The current mainstream three-stage auditing scheme has many significant defects in the context of high concurrency and rapid evolution of content. In terms of semantic understanding, this scheme can only identify explicit violations and is difficult to effectively identify implicit and variant violations such as homophonic, disassembled words, and cross-modal inconsistencies between text and pictures. Rule updates are heavily dependent on expert manual sorting and offline full training, and the average online period for new violation types is not less than 7 days, which cannot respond to sudden public opinion and the problem of rigid rules is prominent. The auditing instances use a fixed number of copies, resulting in an average GPU (Graphics Processing Unit) utilization rate of less than 35%, and resource waste is serious. In addition, existing related technologies are mostly for single modal and require preloading of a large number of regular or artificial features, and cannot solve the problems of instance flexibility and explainability.

[0004] Therefore, how to improve the efficiency, accuracy and flexibility of the auditing operation of user generated content is a technical problem to be solved at present. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a user generated content auditing method and device based on a large model, which can improve the efficiency, accuracy and flexibility of the auditing operation of user generated content. The specific scheme is as follows:

[0006] In a first aspect, the present application provides a user generated content auditing method based on a large model, applied to a system created based on a terminal SDK, an edge node and a center cloud, comprising:

[0007] Obtain multi-modal data of a target instance on a target platform through the terminal SDK, and process the multi-modal data through the edge node to transmit the obtained processed data to the center cloud; the multi-modal data is multi-modal data related to user generated content;

[0008] In the center cloud, a triple is determined based on a target feature of the processed data by a large model, and the processed data with a violation confidence in a preset confidence interval is determined as a target sample;

[0009] In the center cloud, the triple is converted into a resource description framework graph, and if the resource description framework graph matches a rule in a preset rule knowledge base, corresponding violation logic is generated; the preset rule knowledge base is a rule knowledge base in the form of a distributed graph database, which stores rules that violate the target platform;

[0010] In the preset security sandbox of the center cloud, the target sample is audited based on the violation logic by a preset multi-modal model to obtain a corresponding audit result, and the audit result is transmitted to the target platform.

[0011] Optionally, the multi-modal data of the target instance on the target platform is obtained through the terminal SDK, and the multi-modal data is processed through the edge node to transmit the obtained processed data to the center cloud, comprising:

[0012] The video file, the title text and the barrage stream corresponding to the video file are collected through the terminal SDK, and the video file, the title text and the barrage stream are uniformly packaged as multi-modal data in a preset data exchange format, so that the multi-modal data is transmitted to the edge node through a preset two-way secure transmission layer protocol;

[0013] The multi-modal data is processed based on automatic speech recognition technology, optical character recognition technology and a preset video first frame extraction operation within a preset time period through the edge node to obtain processed data, and the processed data is transmitted to the center cloud.

[0014] Optionally, the triple is determined based on a target feature of the processed data by a large model, comprising:

[0015] The target feature of the processed data is obtained by analyzing the processed data through the transformer of the large model and a preset context capacity, and the triple is determined based on the target feature; the preset context capacity is the range of data that can be analyzed by the large model.

[0016] Optionally, if the resource description framework graph matches a rule in a preset rule knowledge base, corresponding violation logic is generated, comprising:

[0017] A corresponding SPARQL query statement is created based on the resource description framework graph, and the SPARQL query statement is used to query in a preset rule knowledge base to obtain a corresponding query result;

[0018] If the query result indicates that the resource description framework graph matches the rules in the preset rule knowledge base, corresponding violation logic is generated.

[0019] Optionally, in the preset security sandbox of the center cloud, the target sample is audited based on the violation logic through a preset multi-modal model to obtain a corresponding audit result, including:

[0020] In the center cloud, a preset security sandbox is determined based on a preset extended Berkeley packet filter, a preset security calculation mode, and a preset security sandbox mechanism.

[0021] In the preset security sandbox, the target sample is audited based on the violation logic through a preset multi-modal model to obtain a corresponding audit result.

[0022] Optionally, the corresponding audit result further includes:

[0023] A preset number of audit results are obtained through a distributed message queue, and the preset number of audit results are processed through a preset real-time data stream processing tool, and the obtained processing results are stored in a preset table.

[0024] In a second aspect, the application provides a user-generated content auditing device based on a large model, applied to a system created based on a terminal SDK, an edge node, and a center cloud, including:

[0025] A data transmission module is configured to obtain multi-modal data of a target instance on a target platform through the terminal SDK, and process the multi-modal data through the edge node to transmit the obtained processed data to the center cloud.

[0026] A sample determination module is configured to determine a triple based on target features of the processed data through a large model in the center cloud, and determine the processed data with a violation confidence in a preset confidence interval as a target sample.

[0027] A logic generation module is configured to convert the triple into a resource description framework graph in the center cloud, and generate corresponding violation logic if the resource description framework graph matches the rules in the preset rule knowledge base. The preset rule knowledge base is a rule knowledge base that stores rules that violate the rules of the target platform in the form of a distributed graph database.

[0028] A result transmission module is configured to audit the target sample based on the violation logic through a preset multi-modal model in a preset security sandbox of the center cloud to obtain a corresponding audit result, and transmit the audit result to the target platform.

[0029] In a third aspect, the present application provides an electronic device, comprising:

[0030] a memory for storing a computer program;

[0031] a processor for executing the computer program to implement the foregoing large model-based user-generated content review method.

[0032] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program; wherein the computer program is executed by a processor to implement the foregoing large model-based user-generated content review method.

[0033] In the present application, the terminal SDK is used to obtain multi-modal data of a target instance on a target platform, and the edge node is used to process the multi-modal data to transmit the obtained processed data to the center cloud; the multi-modal data is multi-modal data related to user-generated content; in the center cloud, a triple is determined based on a target feature of the processed data by a large model, and the processed data with a violation confidence in a preset confidence interval is determined as a target sample; in the center cloud, the triple is converted into a resource description framework graph, and if the resource description framework graph matches a rule in a preset rule knowledge base, corresponding violation logic is generated; wherein the preset rule knowledge base is a rule knowledge base that stores rules that violate the rules of the target platform in the form of a distributed graph database; in the preset security sandbox of the center cloud, the target sample is reviewed based on the violation logic by a preset multi-modal model to obtain a corresponding review result, and the review result is transmitted to the target platform. As can be seen from the above, the present application uses a terminal SDK to obtain multi-modal data of a target instance on a target platform, which is related to user-generated content; the edge node processes these multi-modal data, and then transmits the processed data to the center cloud. In the center cloud, a triple is determined based on a target feature of the processed data by a large model, and the processed data with a violation confidence in a preset confidence interval is determined as a target sample; the triple is converted into a resource description framework graph, and if the graph matches a rule in a preset rule knowledge base, corresponding violation logic is generated, wherein the preset rule knowledge base is a rule knowledge base that stores rules that violate the rules of the target platform in the form of a distributed graph database. Finally, in the preset security sandbox of the center cloud, the target sample is reviewed based on the violation logic by a preset multi-modal model to obtain a review result and transmit it to the target platform. In this way, the present application can improve the efficiency, accuracy and flexibility of the review operation of user-generated content, thereby improving the user experience to a certain extent. BRIEF DESCRIPTION OF DRAWINGS

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0035] Figure 1 This application discloses a flowchart of a user-generated content moderation method based on a large model.

[0036] Figure 2 This application discloses a system architecture diagram for user-generated content moderation based on a large model.

[0037] Figure 3 This is a schematic diagram of a user-generated content moderation device based on a large model disclosed in this application;

[0038] Figure 4 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0039] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Currently, existing user-generated content (UGC) moderation solutions exhibit numerous significant shortcomings in high-concurrency, rapidly evolving content scenarios. In terms of semantic understanding, these solutions can only identify explicit violations, failing to effectively recognize implicit and variant violations such as homophones, word splitting, and cross-modal text-image discrepancies. Rule updates heavily rely on expert manual compilation and offline full-scale training, with the average deployment cycle for new violation types exceeding 7 days, making it unable to cope with sudden public opinion crises and highlighting rule rigidity. The use of a fixed number of review instances results in an average GPU utilization rate of less than 35%, leading to severe resource waste. Furthermore, existing related technologies are mostly targeted at single modalities and require a large number of pre-defined regular expressions or manual features, failing to address the issues of instance flexibility and interpretability. Therefore, this application provides a user-generated content moderation method, apparatus, device, and medium based on a large model, which can improve the efficiency, accuracy, and flexibility of user-generated content moderation operations.

[0041] See Figure 1 and Figure 2 As shown, this invention discloses a user-generated content moderation method based on a large model, applied to a system created based on terminal SDK, edge nodes, and a central cloud, including:

[0042] Step S11: Obtain multimodal data of the target instance on the target platform through the terminal SDK, and process the multimodal data through the edge node to transmit the processed data to the central cloud; the multimodal data is multimodal data related to user-generated content.

[0043] In this embodiment, the terminal SDK (Software Development Kit) first obtains multimodal data related to the user-generated content corresponding to the target instance from the target platform, then transmits it to the edge node for processing, and the processed data will be sent to the central cloud.

[0044] Specifically, firstly, the terminal SDK collects video files related to the target instance on the target platform, along with corresponding title text and bullet comment streams. The video files, title text, and bullet comment streams are then uniformly encapsulated into multimodal data in a preset data exchange format. This multimodal data is then transmitted to the edge node via a preset bidirectional secure transport layer protocol. The preset data exchange format can be ProtocolBuffer; the preset bidirectional secure transport layer protocol can be mTLS (Mutual Transport Layer Security). It should be noted that the aforementioned target instance can refer to content carriers created or uploaded by users on the target platform; it can be a single content unit or a collection of related content.

[0045] Furthermore, the edge nodes process the multimodal data within a preset time period using Automatic Speech Recognition (ASR), Optical Character Recognition (OCR), and a preset first-frame video extraction operation to obtain processed data, which is then transmitted to the central cloud. The preset time period can be 20ms.

[0046] Step S12: In the central cloud, the triples are determined based on the target features of the processed data by the large model, and the processed data with violation confidence levels in the preset confidence interval are determined as target samples.

[0047] In this embodiment, after receiving processed data transmitted from edge nodes, the central cloud analyzes the processed data using a large language model (LLM) transformer and a preset context capacity to obtain the target features of the processed data and determine triples based on these target features. In other words, when analyzing the processed data, the large model first extracts and encodes features using a transformer, while simultaneously using a preset context capacity limit to ensure the analysis process is within the effective processing range of the large model. The preset context capacity can be dynamically adjusted based on the average length of the processed data to balance analysis accuracy and efficiency. The preset context capacity represents the range of data that the large model can analyze.

[0048] Furthermore, the calculation of the violation confidence level for the processed data can be based on the identification results of violation features in the processed data by a large model. When the confidence level falls within a pre-set confidence interval, it is marked as a target sample requiring further review. The pre-set confidence interval can be 0.7-0.9.

[0049] Step S13: In the central cloud, the triple is transformed into a resource description framework graph. If the resource description framework graph matches the rules in the preset rule knowledge base, the corresponding violation logic is generated. The preset rule knowledge base is a rule knowledge base that stores rules that violate the target platform in the form of a distributed graph database.

[0050] In this embodiment, after the central cloud transforms the triples into a Resource Description Framework (RDF) graph, it creates a corresponding SPARQL query statement based on the RDF graph and uses the SPARQL query statement to perform a query in a preset rule knowledge base to obtain the corresponding query results. SPARQL is a standardized query language specifically used for querying and manipulating RDF data.

[0051] If the query result shows that the resource description framework diagram matches a rule in the preset rule knowledge base, then corresponding violation logic is generated. It can be understood that the process of querying the preset rule knowledge base using SPARQL queries is essentially the process of matching the resource description framework diagram with the rules in the preset rule knowledge base. During the matching process, the SPARQL query statement created based on the resource description framework diagram will accurately locate the corresponding rule entry in the knowledge base. Furthermore, the violation logic may include violation judgment criteria, associated feature weights, etc.

[0052] Step S14: In the preset security sandbox of the central cloud, the target sample is reviewed based on the violation logic using a preset multimodal model to obtain the corresponding review result, and the review result is transmitted to the target platform.

[0053] In this embodiment, within the preset security sandbox of the central cloud, a preset multimodal model combined with violation logic is used to review the target sample, generate review results, and feed them back to the target platform.

[0054] Specifically, regarding the architecture of the preset security sandbox, in the central cloud, the preset security sandbox is determined based on a preset extended Berkeley Packet Filter (eBPF), a preset secure computing mode (SecureComputing Mode), and a preset security sandbox mechanism (Landlock). The preset extended Berkeley Packet Filter can be used to control data interaction between the sandbox and the outside world, allowing only necessary data flows to pass through; the preset secure computing mode can employ hardware-level isolation technology to ensure that data is not accessed without authorization during the audit process.

[0055] Within the preset security sandbox, the target sample is reviewed using a preset multimodal model based on the violation logic, yielding corresponding review results. That is, during the review process, the preset multimodal model cross-validates the multimodal features of the target sample according to the judgment criteria in the violation logic.

[0056] After the review results are generated, a preset number of review results are retrieved through a distributed message queue (i.e., Kafka). These preset number of review results are then processed using a preset real-time data stream processing tool. The processed results are stored in a preset table, which facilitates querying and retrieval by the target platform and provides data support for subsequent iterative optimization of the rule knowledge base. The preset real-time data stream processing tool can be Spark Streaming; the preset table can be Iceberg.

[0057] Furthermore, in this embodiment, real-time metric monitoring enables elastic resource scheduling and model performance iteration. The dynamic instance management engine in the system continuously collects system operation metrics, including queries per second (QPS), quantile response time, and GPU utilization. When the monitored data meets preset expansion conditions, a horizontal expansion process is automatically triggered. This means that 5% of the traffic can be redirected to a preset number of newly deployed compute instances. Once the readiness probe verifies that the model loading time is less than a preset number of seconds, the new instances officially take over the business traffic, ensuring uninterrupted audit services during the expansion process. After the traffic peak, if GPU utilization remains below a preset utilization threshold for a certain period, the system automatically performs a scaling-down operation, adjusting the number of instances accordingly to achieve a balance between efficient resource utilization and service performance.

[0058] Furthermore, incremental fine-tuning of the model in the system can be performed within a specified time period, employing Low-Rank Adaptation (LoRA) technology. This involves setting rank parameters and scaling factors, and completing model updates for a certain duration on a single GPU. Differential privacy protection mechanisms are enabled during fine-tuning to ensure data security while optimizing the model. Moreover, performance can be replayed using validation data. If the model's F1 score (a metric used to evaluate the performance of binary classification models) improves to a preset threshold, a rollback mechanism is not required.

[0059] In addition, the monitoring component in the system (i.e., Prometheus) can monitor fluctuations in traffic replay metrics at the first preset time and check whether the metric reaches a preset alarm threshold. If the preset alarm threshold is reached, corresponding alarm information is generated. Furthermore, when the elastic scheduling function needs to be suspended at the second preset time, after the operation and maintenance command is triggered, the dynamic instance management engine first gradually reduces the number of instance replicas to 0 to ensure a smooth termination of the service, and then deletes the custom resources of the inference set to achieve complete reclamation of GPU resources. The entire process leaves no residual resource occupation, ensuring clean management of system resources.

[0060] As can be seen from the above, this application utilizes the terminal SDK to obtain multimodal data of target instances on the target platform, which is related to user-generated content. Edge nodes process this multimodal data, and then the processed data is transmitted to the central cloud. In the central cloud, a large model is used to determine triples based on the target features of the processed data. Processed data with violation confidence levels within a pre-set confidence interval are designated as target samples. The triples are transformed into a resource description framework graph. If this graph matches a rule in a pre-set rule knowledge base, corresponding violation logic is generated. This pre-set rule knowledge base is stored in a distributed graph database and contains rule knowledge bases that violate the target platform's rules. Finally, in the central cloud's pre-set security sandbox, the target samples are reviewed based on the violation logic using a pre-set multimodal model, and the review results are transmitted to the target platform. In this way, this application can improve the efficiency, accuracy, and flexibility of reviewing user-generated content, thereby improving the user experience to a certain extent.

[0061] Accordingly, see Figure 3 As shown, this application provides a user-generated content moderation device based on a large model, applied to a system created based on terminal SDK, edge nodes, and central cloud, including:

[0062] The data transmission module 11 is used to obtain multimodal data of the target instance on the target platform through the terminal SDK, and process the multimodal data through the edge node to transmit the processed data to the central cloud.

[0063] The sample determination module 12 is used to determine triples based on the target features of the processed data in the central cloud using a large model, and to determine the processed data with violation confidence in a preset confidence interval as target samples.

[0064] The logic generation module 13 is used to convert the triples into a resource description framework diagram in the central cloud. If the resource description framework diagram matches the rules in the preset rule knowledge base, the corresponding violation logic is generated. The preset rule knowledge base is a rule knowledge base that stores rules that violate the target platform in the form of a distributed graph database.

[0065] The result transmission module 14 is used to review the target sample based on the violation logic using a preset multimodal model in the preset security sandbox of the central cloud, obtain the corresponding review result, and transmit the review result to the target platform.

[0066] As can be seen from the above, this application utilizes the terminal SDK to obtain multimodal data of target instances on the target platform, which is related to user-generated content. Edge nodes process this multimodal data, and then the processed data is transmitted to the central cloud. In the central cloud, a large model is used to determine triples based on the target features of the processed data. Processed data with violation confidence levels within a pre-set confidence interval are designated as target samples. The triples are transformed into a resource description framework graph. If this graph matches a rule in a pre-set rule knowledge base, corresponding violation logic is generated. This pre-set rule knowledge base is stored in a distributed graph database and contains rule knowledge bases that violate the target platform's rules. Finally, in the central cloud's pre-set security sandbox, the target samples are reviewed based on the violation logic using a pre-set multimodal model, and the review results are transmitted to the target platform. In this way, this application can improve the efficiency, accuracy, and flexibility of reviewing user-generated content, thereby improving the user experience to a certain extent.

[0067] In some specific embodiments, the data transmission module 11 specifically includes:

[0068] The first data transmission unit is used to collect video files about the target instance and title text and bullet comments corresponding to the video files on the target platform through the terminal SDK, and encapsulate the video files, title text and bullet comments into multimodal data in a preset data exchange format, so as to transmit the multimodal data to the edge node through a preset bidirectional secure transmission layer protocol;

[0069] The second data transmission unit is used to process the multimodal data through the edge node within a preset time period based on automatic speech recognition technology, optical character recognition technology, and preset video first frame extraction operation to obtain processed data, and then transmit the processed data to the central cloud.

[0070] In some specific embodiments, the sample determination module 12 specifically includes:

[0071] The triplet determination unit is used to analyze the processed data through the transformer of a large model and a preset context capacity to obtain the target features of the processed data, and to determine triples based on the target features; wherein, the preset context capacity is the range of data that the large model can analyze.

[0072] In some specific embodiments, the logic generation module 13 specifically includes:

[0073] The rule query unit is used to create a corresponding SPARQL query statement based on the resource description framework diagram, and use the SPARQL query statement to query in a preset rule knowledge base to obtain the corresponding query results;

[0074] The logic generation unit is used to generate corresponding violation logic if the query result shows that the resource description framework diagram matches the rules in the preset rule knowledge base.

[0075] In some specific embodiments, the result transmission module 14 specifically includes:

[0076] The sandbox determination unit is used to determine a preset security sandbox in the central cloud based on a preset extended Berkeley packet filter, a preset secure computing mode, and a preset security sandbox mechanism.

[0077] The sample review unit is used to review the target sample based on the violation logic using a preset multimodal model in the preset security sandbox, and obtain the corresponding review results.

[0078] In some specific embodiments, the result transmission module 14 further includes:

[0079] The result storage unit is used to obtain a preset number of review results through a distributed message queue, process the preset number of review results through a preset real-time data stream processing tool, and store the processed results in a preset table.

[0080] Furthermore, embodiments of this application also disclose an electronic device, Figure 4 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the user-generated content moderation method based on a large model disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be a computer.

[0081] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0082] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0083] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including computer programs capable of performing the large-model-based user-generated content moderation method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0084] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned user-generated content moderation method based on a large model. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0086] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0087] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0088] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0089] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for reviewing user-generated content based on a large model, characterized in that, Applicable to systems built on terminal SDKs, edge nodes, and central clouds, including: The terminal SDK is used to obtain multimodal data of the target instance on the target platform, and the multimodal data is processed by the edge node to transmit the processed data to the central cloud; the multimodal data is multimodal data related to user-generated content. In the central cloud, triples are determined based on the target features of the processed data using a large model, and the processed data with violation confidence levels within a preset confidence interval are identified as target samples. In the central cloud, the triples are transformed into resource description framework graphs. If the resource description framework graph matches a rule in a preset rule knowledge base, a corresponding violation logic is generated. The preset rule knowledge base is a rule knowledge base that stores rules that violate the target platform in the form of a distributed graph database. In the preset security sandbox of the central cloud, the target sample is reviewed based on the violation logic using a preset multimodal model to obtain the corresponding review results, and the review results are transmitted to the target platform.

2. The user-generated content review method based on a large model according to claim 1, characterized in that, The step of acquiring multimodal data of a target instance on the target platform through the terminal SDK, processing the multimodal data through the edge node, and transmitting the processed data to the central cloud includes: The terminal SDK collects video files of the target instance on the target platform, as well as the title text and bullet comments corresponding to the video files. The video files, title text, and bullet comments are then uniformly encapsulated into multimodal data in a preset data exchange format. The multimodal data is then transmitted to the edge node through a preset bidirectional secure transport layer protocol. The edge nodes process the multimodal data within a preset time period using automatic speech recognition technology, optical character recognition technology, and preset video first frame extraction operations to obtain processed data, which is then transmitted to the central cloud.

3. The user-generated content review method based on a large model according to claim 1, characterized in that, The process of determining triples based on target features of the processed data using a large model includes: The processed data is analyzed using a large model transformer and a preset context capacity to obtain target features of the processed data, and triples are determined based on the target features; wherein, the preset context capacity is the range of data that the large model can analyze.

4. The user-generated content review method based on a large model according to claim 1, characterized in that, If the resource description framework diagram matches a rule in the preset rule knowledge base, then corresponding violation logic is generated, including: Based on the resource description framework diagram, a corresponding SPARQL query statement is created, and the SPARQL query statement is used to perform a query in a preset rule knowledge base to obtain the corresponding query results. If the query results show that the resource description framework diagram matches the rules in the preset rule knowledge base, then the corresponding violation logic is generated.

5. The user-generated content review method based on a large model according to claim 1, characterized in that, In the preset security sandbox of the central cloud, the target sample is reviewed based on the violation logic using a preset multimodal model to obtain the corresponding review results, including: In the central cloud, a preset security sandbox is determined based on a preset extended Berkeley packet filter, a preset secure computing mode, and a preset security sandbox mechanism; In the preset security sandbox, the target sample is reviewed based on the violation logic using a preset multimodal model to obtain the corresponding review results.

6. The user-generated content moderation method based on a large model according to any one of claims 1 to 5, characterized in that, The process of obtaining the corresponding review results also includes: A preset number of review results are obtained through a distributed message queue, and the preset number of review results are processed through a preset real-time data stream processing tool. The processed results are then stored in a preset table.

7. A user-generated content moderation device based on a large model, characterized in that, Applicable to systems built on terminal SDKs, edge nodes, and central clouds, including: The data transmission module is used to obtain multimodal data of the target instance on the target platform through the terminal SDK, and process the multimodal data through the edge node to transmit the processed data to the central cloud. The sample determination module is used to determine triples based on the target features of the processed data using a large model in the central cloud, and to determine the processed data whose violation confidence is within a preset confidence interval as target samples. The logic generation module is used to transform the triples into a resource description framework diagram in the central cloud. If the resource description framework diagram matches the rules in the preset rule knowledge base, the corresponding violation logic is generated. The preset rule knowledge base is a rule knowledge base that stores rules that violate the target platform in the form of a distributed graph database. The result transmission module is used to review the target sample based on the violation logic using a preset multimodal model in the preset security sandbox of the central cloud, obtain the corresponding review result, and transmit the review result to the target platform.

8. The user-generated content review device based on a large model according to claim 7, characterized in that, The result transmission module includes: The sandbox determination unit is used to determine the preset security sandbox based on the preset extended Berkeley package filter, the preset secure computing mode, and the preset security sandbox mechanism. The sample review unit is used to review the target sample based on the violation logic using a preset multimodal model in the preset security sandbox, and obtain the corresponding review results.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the user-generated content moderation method based on a large model as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the user-generated content moderation method based on a large model as described in any one of claims 1 to 6.

Images