Information dissemination methods, devices, equipment, and media based on multi-agent collaboration

By employing a multi-agent collaborative information publishing method, and utilizing semantic parsing and a two-layer sandbox architecture to generate and review content, the efficiency and security issues of information publishing on multimedia platforms are resolved, achieving efficient and accurate content publishing.

CN123115535APending Publication Date: 2026-11-17ZHILONG INNOVATION (BEIJING) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610764485.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for information dissemination on multimedia platforms suffer from high costs and long processing times in generating prompts, the generation of content may contain sensitive information, and the dissemination process is cumbersome and time-consuming, making it difficult to adapt to large-scale content dissemination.

Method used

A multi-agent collaborative information publishing method is adopted. The first agent performs semantic content parsing and generates skill module calls, and uses a two-layer sandbox architecture to generate the initial publishing content. The second agent reviews the content, the third agent makes adjustments, and finally the content is published using a temporary token mechanism.

Benefits of technology

It enables efficient and accurate information dissemination, shortens generation and review time, reduces the risk of sensitive content, improves dissemination efficiency, and is suitable for large-scale content dissemination.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115535A_ABST
    Figure CN123115535A_ABST
Patent Text Reader

Abstract

This disclosure presents an information publishing method, apparatus, device, and medium based on multi-agent collaboration. One specific implementation of the method includes: in response to selecting generation skill module information and obtaining an information publishing instruction, executing a first generation step: performing semantic content parsing on the information publishing instruction to obtain semantic parsing information; in response to the requirement that the information representation time interval of the published content is less than a target duration, sequentially calling each generation skill module; for the initial published content, executing a first processing step: sequentially calling each review skill module to obtain a review result; in response to the presence of review issues in the representation, performing content adjustment to obtain the published content; in response to the review terminal's feedback that the representation is correct, publishing the published content to the publishing platform. This implementation, utilizing multiple agents, can efficiently and accurately automatically publish the generated published content to the relevant platform after inputting the corresponding information publishing instruction on the information processing platform.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of intelligent agent technology, and more specifically to information dissemination methods, apparatuses, electronic devices, and computer-readable media based on multi-agent collaboration. Background Technology

[0002] Currently, with the development of multimedia platforms, information dissemination on these platforms has become one of the main methods of information dissemination. The typical approach to generating information for multimedia platforms is as follows: First, obtain the basic dataset used for content generation. Then, based on a large model using prompt word technology, generate the content corresponding to the basic dataset. Finally, upload the reviewed content to the corresponding publishing platform.

[0003] However, when using the above method, the following technical problems often arise: Complex and precise prompts need to be set based on a basic dataset, which is costly and time-consuming. Furthermore, the output of large models often contains sensitive content (e.g., hallucinations and copyright infringement), resulting in significant flaws in the published content. This necessitates multiple rounds of prompt generation and refinement, which is also time-consuming. Finally, the generation, uploading, and review of published content all require actual page processing and review operations, making the overall publishing process cumbersome and time-consuming, unsuitable for large-scale content publishing.

[0004] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0006] Some embodiments of this disclosure propose information publishing methods, apparatuses, electronic devices, and computer-readable media based on multi-agent collaboration to solve one or more of the technical problems mentioned in the background section above.

[0007] In a first aspect, some embodiments of this disclosure provide an information publishing method based on multi-agent collaboration, comprising: in response to selecting generation skill module information on an information processing platform and obtaining an input information publishing instruction, using a first agent, performing the following first generation step: performing semantic content parsing on the information publishing instruction to obtain semantic parsing information, wherein the semantic parsing information includes: publishing platform and publishing content requirement information; in response to the publishing content requirement information indicating that the time interval between the publishing time and the current time is less than the target duration, according to the semantic parsing information, using a first two-layer sandbox architecture, sequentially calling each generation skill module in the generation skill module information to generate initial publishing content under the publishing platform; for the initial publishing... The content is published, and a second intelligent agent is used to execute the first processing step: obtain the review skill module information corresponding to the aforementioned publishing platform; use a second two-layer sandbox architecture to sequentially call each review skill module in the aforementioned review skill module information to review the content of the initial published content and obtain the review result; in response to the aforementioned review result indicating that there is a review problem, use a third intelligent agent bound to the aforementioned information processing platform to generate adjusted content for the aforementioned initial published content; use the aforementioned first intelligent agent and / or the aforementioned second intelligent agent to execute the aforementioned adjusted content and obtain the published content; in response to receiving the review terminal feedback result for the aforementioned published content and the review terminal feedback result indicating that it is correct, use a temporary token mechanism to publish the aforementioned published content to the aforementioned publishing platform.

[0008] Secondly, some embodiments of this disclosure provide an information publishing device based on multi-agent collaboration, comprising: a first execution unit configured to, in response to selecting generation skill module information on an information processing platform and obtaining an input information publishing instruction, utilize a first agent to execute the following first generation step: performing semantic content parsing on the information publishing instruction to obtain semantic parsing information, wherein the semantic parsing information includes: publishing platform and publishing content requirement information; in response to the publishing content requirement information indicating that the time interval between the publishing time and the current time is less than a target duration, according to the semantic parsing information, using a first two-layer sandbox architecture, sequentially calling each generation skill module in the generation skill module information to generate initial publishing content under the publishing platform; and a second execution unit configured to, for the initial publishing content... The system utilizes a second intelligent agent to execute the first processing step: obtaining the review skill module information corresponding to the aforementioned publishing platform; using a second two-layer sandbox architecture, sequentially calling each review skill module in the aforementioned review skill module information to review the content of the initial published content and obtain the review result; a generation unit is configured to, in response to the aforementioned review result indicating the existence of review issues, use the third intelligent agent bound to the aforementioned information processing platform to generate adjusted content for the aforementioned initial published content; a third execution unit is configured to, using the aforementioned first intelligent agent and / or the aforementioned second intelligent agent, execute the aforementioned adjusted content to obtain the published content; and a publishing unit is configured to, in response to receiving the review terminal feedback result for the aforementioned published content and the review terminal feedback result indicating that it is correct, use a temporary token mechanism to publish the aforementioned published content to the aforementioned publishing platform.

[0009] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, such that when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0010] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method as described in any implementation of the first aspect.

[0011] The above-described embodiments of this disclosure have the following beneficial effects: Through the information publishing method based on multi-agent collaboration in some embodiments of this disclosure, multiple agents can efficiently and accurately publish the generated content to the relevant platform after inputting the corresponding information publishing instruction on the information processing platform. Specifically, the inefficiency of related content publishing stems from the following: complex and precise prompts need to be set based on a basic dataset, resulting in high cost and time consumption for prompt generation. Furthermore, the content output by large models contains a significant amount of sensitive content (e.g., hallucination content and copyright-infringing content), leading to substantial defects in the published content. Multiple refinements of prompts are required, which is time-consuming. Finally, the generation, uploading, and review of published content all require actual execution of related page processing and review operations, making the overall content publishing process cumbersome and time-consuming, unsuitable for large-scale content publishing. Based on this, some embodiments of the information publishing method based on multi-agent collaboration disclosed herein firstly, in response to selecting generation skill module information on the information processing platform and obtaining the input information publishing instruction, the first agent performs the following first generation step: Step 1: Semantic content parsing of the aforementioned information publishing instruction yields semantic parsing information, which includes publishing platform and publishing content requirement information. Here, the information processing platform is set up so that skill module information and information publishing instructions can be selected on the information processing platform to accurately and efficiently generate the required publishing content based on the information publishing requirements and the content generation task module. Based on the first agent, by utilizing the generation skill module information, various logical modules for generating publishing content are automatically invoked, enabling efficient and streamlined generation of publishing content. By performing semantic content parsing on the information publishing instruction, the content requirements and publishing platform information used for publishing content generation can be obtained. Step 2: In response to the aforementioned publishing content requirement information indicating that the time interval between the publishing time and the current time is less than the target duration, based on the aforementioned semantic parsing information, each generation skill module in the aforementioned generation skill module information is sequentially invoked using a first two-layer sandbox architecture to generate the initial publishing content under the aforementioned publishing platform. Here, when the required information representation for the published content is shorter than the target duration, a comprehensive processing approach can be adopted: first, an initial published content is generated by a first intelligent agent, and then a second intelligent agent reviews the initial published content. This significantly reduces the efficiency of both generation and review. Furthermore, utilizing the first two-layer sandbox architecture avoids permission violations during the invocation of various generation skill modules, preventing financial losses. By sequentially executing each generation skill module, the initial published content for the publishing platform can be generated automatically and accurately.Then, for the aforementioned initial published content, the second intelligent agent executes the first processing step: First, it obtains the review skill module information corresponding to the aforementioned publishing platform, so as to automatically and accurately review the initial published content based on each review skill module, thereby ensuring the quality of the initial published content. Second, using the second two-layer sandbox architecture, it sequentially calls each review skill module in the aforementioned review skill module information to review the content of the aforementioned initial published content and obtain the review result. Here, the use of the second two-layer sandbox architecture can avoid the problem of unauthorized access and avoid loss of property. Moreover, using each review skill module can accurately and efficiently perform semantic review of the content of the initial published content. Next, in response to the aforementioned review result indicating the existence of review issues, the third intelligent agent bound to the aforementioned information processing platform generates adjusted content for the aforementioned initial published content. Here, the third intelligent agent can make timely adjustments to the content with review issues to assist in the accurate generation of high-quality published content subsequently. Then, using the aforementioned first intelligent agent and / or the aforementioned second intelligent agent, the aforementioned adjusted content is executed to obtain high-quality published content. Finally, in response to receiving feedback from the review terminal regarding the aforementioned published content, and confirming that the feedback is accurate, the published content is then published to the aforementioned publishing platform using a temporary token mechanism. This temporary token mechanism, through its time-limited expiration and least privilege design, significantly reduces the risks of account theft and data leakage. In summary, through various intelligent agents with different functions, after inputting corresponding information publishing commands on the information processing platform, the generated content can be automatically and efficiently published to the relevant platforms. Attached Figure Description

[0012] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0013] Figure 1 This is a flowchart of some embodiments of the information dissemination method based on multi-agent collaboration according to the present disclosure; Figure 2 These are schematic diagrams of some embodiments of the information publishing device based on multi-agent collaboration according to this disclosure; Figure 3 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0015] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0016] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0017] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0018] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0019] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] refer to Figure 1 The diagram illustrates a flow 100 of some embodiments of a multi-agent cooperative information publishing method according to the present disclosure. This multi-agent cooperative information publishing method includes the following steps: Step 101: In response to selecting the skill module information to be generated on the information processing platform and obtaining the input information publishing instruction, the first intelligent agent executes the following first generation step: Step 1011: Perform semantic content parsing on the above information release instruction to obtain semantic parsing information.

[0021] In some embodiments, the executing entity (e.g., an electronic device) of the above-mentioned information publishing method based on multi-agent collaboration can perform semantic content parsing on the above-mentioned information publishing instructions to obtain semantic parsing information. The information processing platform can be an intelligent platform supporting various information processing functions. In practice, the information processing platform can be a platform deploying diverse intelligent agents. In practice, these various information processing functions can include: platform account creation, content information publishing, content information generation, content information customization, AI video generation, and AI image and text generation. In practice, the information processing platform can be a platform built on OpenClaw. The homepage of the information processing platform displays a content publishing control. Clicking the content publishing control redirects to a page that supports content publishing (i.e., the content publishing generation page). The content publishing page displays an account login area, a content publishing generation area, and a skill module selection area. The account login area can be the area where the platform account information for logging in and publishing content is displayed. The content publishing generation area can be the page area for generating published content. In practice, the content publishing generation area includes an information publishing instruction input control and a content display area. The information publishing instruction corresponding to the content to be generated can be entered in the instruction input control. During the content generation and display process, the content generated by the agent and the corresponding review results will be displayed in real time in the content display area. The generation skill module information can be module information for each generation skill module used to subsequently generate and publish content. For example, module information can be a module identifier. Generation skill modules can be task skill modules used to generate and publish content. Task skill modules are extensible functional modules in the OpenClaw framework, essentially standardized "operation manuals + execution toolkits." Task skill modules endow the agent with specific execution capabilities (e.g., operating a browser, processing files, calling APIs), upgrading the large model from "only able to chat" to a digital employee capable of proactively completing actual tasks. Specifically, in the content publishing scenario, each generation skill module can be a skill module corresponding to each process node in the content generation process. That is, each process node has a corresponding skill module. Skill modules can be the execution content of the corresponding process node. For example, each generation skill module can include: a hot data acquisition module and a copywriting generation module. Here, the hot data acquisition module can be a skill module that acquires a hot dataset at the current time. The copywriting generation module can be a skill module that generates corresponding copywriting based on the hot dataset. Hot topic data can be news data whose current discussion volume exceeds the target value. The various generation skill modules in the generation skill module information can be pre-set. That is, in the generation skill module selection area, users can select all the generation skill modules that support generating and publishing content. Users publishing content can customize the selection of corresponding generation skill modules in the generation skill module selection area.Among the various generated skill modules displayed, several must be selected (e.g., the copywriting generation module). Information publishing instructions can be user commands to publish information. These instructions can be input by the user publishing content. For example, an information publishing instruction could be "Publish the content corresponding to Hot News A on account 9***72". It should be noted that during the input of information publishing instructions, the input control can display the necessary elements in real time to ensure the completeness of the instructions and prevent the subsequent output content from not meeting user expectations. The first intelligent agent can be an intelligent agent deployed by the information processing platform specifically for generating published content. The first intelligent agent has the following functions: specialized design in vertical domains, fully automated workflow, multimodal content generation capabilities, and memory and decision optimization mechanisms. Specialized design in vertical domains can be the construction of dedicated workflows for different content scenarios (e.g., short video marketing, news reporting, e-commerce copywriting), with built-in industry terminology libraries, compliance rule libraries, and historical case libraries. Fully automated workflows can be the fully automated processing of published content generation based on each generated skill module. Multimodal content generation capabilities can support chained production of "text -> image -> video" and optimize output formats for different channels. The memory and decision-making optimization mechanism can store historical publishing data (such as click-through rates for certain types of titles and user interaction preferences) for optimizing subsequent strategies. It can also track the entire content chain from publishing to conversion, automatically attributing key influencing factors (e.g., the impact of cover image style on completion rate). The first intelligent agent can be deployed based on a private large model, a local knowledge base, and a security gateway. The large model corresponding to the first intelligent agent can be a multimodal large model (which can be a large model with a conventional architecture trained privately). Semantic content parsing can extract key publishing information from user-published content. The publishing platform can be the platform for subsequently generated publishing content. Publishing content requirement information can be the requirements for generating publishing content. For example, publishing content requirement information can include: publishing content generation time, publishing content style, and the corresponding content preference.

[0022] As an example, the large model deployed corresponding to the first intelligent agent can be used to perform semantic content parsing on the above information publishing instructions to obtain semantic parsing information.

[0023] In some optional implementations of certain embodiments, after step 1011, the steps further include: The first step is to use the aforementioned third intelligent agent to construct a first two-layer sandbox structure for generating skill module information and a second two-layer sandbox structure for reviewing skill module information.

[0024] As an example, firstly, the aforementioned executing entity can obtain container processing instructions corresponding to the generated skill module information. These container processing instructions can be instructions on how each generated skill module should be executed within a container, customized by relevant technical personnel according to its corresponding skill effects. Specifically, if the relevant technical personnel have not performed customized customization, the third-party intelligent agent can implement container customization for each generated skill module based on historical experience and the corresponding skill effects. The skill effect corresponding to the generated skill module can be the purpose to be achieved by executing the corresponding logical semantics of the generated skill module—that is, the content to be generated. In addition, the container processing instructions also include information on how to trigger manual intervention in cases of unauthorized access during the execution of the generated skill module. Then, using the container processing instructions and the third-party intelligent agent, a first two-layer sandbox structure is constructed for the aforementioned generated skill module information. Furthermore, the third-party intelligent agent determines the information publishing direction and publishing platform information corresponding to the information publishing instructions. Then, based on the information publishing direction and publishing platform information, the third-party intelligent agent determines the container processing arrangements for each review skill module and the manual intervention method in cases of unauthorized access. Finally, based on the container processing arrangements and methods corresponding to each audit skill module, a second two-layer sandbox structure is constructed using a third intelligent agent.

[0025] The second step involves setting up a first permission monitoring thread that runs when the first intelligent agent executes the generated skill module information, and a second permission monitoring thread that runs when the second intelligent agent executes the reviewed skill module information. The first permission monitoring thread can monitor for permission violations when each generated skill module is executed in its corresponding container. The second permission monitoring thread can monitor for permission violations when each reviewed skill module is executed in its corresponding container. In practice, both the first and second permission monitoring threads execute in real-time during the content generation process. That is, they execute in parallel with both the content generation and review processes.

[0026] Step 1012: In response to the above-mentioned requirement that the time interval between the release time and the current time is less than the target duration, based on the above-mentioned semantic parsing information, the first two-layer sandbox architecture is used to sequentially call each generation skill module in the above-mentioned generation skill module information to generate the initial release content under the above-mentioned release platform.

[0027] In some embodiments, in response to the requirement that the time interval between the publication time and the current time is less than the target duration, the executing entity can, based on the semantic parsing information, utilize the first two-layer sandbox architecture to sequentially call each generation skill module in the generation skill module information to generate the initial publication content on the publishing platform. The publication time can be the time when the publication content is published to the publishing platform. The target duration can be a time threshold used to measure whether the publication content needs to be generated quickly. For example, the target duration can be 20 minutes. The first two-layer sandbox architecture can be set for each generation skill module in the generation skill module information. The two-layer sandbox architecture is a specific implementation scheme that prevents the automated operations of the first intelligent agent when executing each generation skill module from causing account bans or data leaks through dual protection of technical isolation and manual intervention. In practice, the first layer can use Docker containers to limit the physical operation scope of the code. The second layer can use manual confirmation to block high-risk operation chains. Each generation skill module has a corresponding execution order. The initial publication content can be the preliminarily generated publication content awaiting further review.

[0028] As an example, based on the first two-layer sandbox architecture and the content corresponding to the semantic parsing information, the first intelligent agent is instructed to execute in sequence according to the execution order of each generated skill module to obtain the initial published content.

[0029] In some optional implementations of certain embodiments, the above-mentioned method of generating skill modules in the above-mentioned skill module information by sequentially calling each generation skill module in the above-mentioned generation skill module information based on the above-mentioned semantic parsing information and utilizing the first two-layer sandbox architecture to generate the initial release content under the above-mentioned release platform includes the following steps: The first step, based on the semantic parsing information described above, utilizes the first two-layer sandbox architecture and the first permission monitoring thread to sequentially call each generation skill module in the generation skill module information to generate the initial release content under the publishing platform. Specifically, if the first permission monitoring thread detects an unauthorized access behavior when the first agent executes the corresponding generation skill module, it directly triggers the first agent to pause processing until it restarts processing on the corresponding permission recovery page of the information processing platform. Further details are omitted.

[0030] The second step involves, based on the initial published content, utilizing the second two-layer sandbox architecture and the second permission monitoring thread, sequentially calling each of the aforementioned review skill modules in the review skill module information to review the initial published content and obtain the review results. Specifically, if the second permission monitoring thread detects an unauthorized access behavior when the second agent executes a corresponding review skill module, it directly triggers the second agent to pause processing until it restarts processing on the corresponding permission recovery page of the information processing platform. Further details are omitted.

[0031] In some optional implementations of certain embodiments, the various generation skill modules include: a data acquisition skill module, a data processing skill module, and a copywriting generation skill module. The data acquisition skill module can be a skill module for acquiring datasets related to subsequently published content. In practice, the published content is generated based on the dataset acquired by the data acquisition skill module. The data processing skill module can be a skill module for performing various processing on the acquired dataset. In practice, these processing methods may include: structuring, desensitization, and error correction. Based on the data processing skill module, structured data results that facilitate subsequent copywriting generation can be obtained. The copywriting generation skill module can be a skill module for generating copywriting content based on the data processing results. Specifically, the data acquisition skill module, data processing skill module, and copywriting generation skill module can all call the large model corresponding to the first large model to perform specific tasks.

[0032] Optionally, based on the semantic parsing information, the first two-layer sandbox architecture is used to sequentially call each generation skill module in the generation skill module information to generate the initial release content under the release platform, including the following steps: The first step is to acquire the basic publishing data within the container corresponding to the data acquisition skill module, based on the aforementioned publishing content requirements. The container corresponding to the data acquisition skill module can be the container of the execution environment corresponding to the data acquisition skill module. That is, the data acquisition operation is performed within this container. The basic publishing data can be the foundational data used for generating the publishing content. In practice, the basic data may include: contextual data and hotspot data corresponding to each key requirement of the publishing content requirements.

[0033] The second step involves responding to the detection during the acquisition process that the relevant data acquisition behavior is in the first abnormal behavior list and / or the aforementioned first permission monitoring thread detects an unauthorized access situation. A data alert pop-up window appears on the content generation page of the aforementioned information processing platform to handle the data alert. The first abnormal behavior can be an abnormal behavior related to data acquisition. Each of the first abnormal behaviors in the first abnormal behavior list is pre-built. When a relevant first abnormal behavior is detected being executed within the corresponding container content, a data alert pop-up window will appear on the content generation page to alert to the existence of the abnormal behavior and to promptly handle the abnormal behavior operation. In practice, the first abnormal behavior list may include: abnormal behaviors corresponding to data source access and abnormal behaviors corresponding to permissions and data flow. Abnormal behaviors corresponding to data source access may include: whether external API calls are abnormal and whether there is unauthorized access to the local file system. Abnormal behaviors corresponding to permissions and data flow may include: whether behaviors exceeding declared permissions are executed (e.g., Skills declares "read-only" but attempts to write to a file), and whether the acquired data is transferred to an unauthorized destination (e.g., an unauthorized destination could be an overseas server or a non-business-related domain).

[0034] The third step involves processing the published basic data within the container corresponding to the aforementioned data processing skill module to obtain the data processing result. This result can be the result of preprocessing the published basic data. The container corresponding to the data processing skill module can be the container of the execution environment corresponding to the data processing skill module; that is, the data processing operation is performed within this container.

[0035] The fourth step involves responding to the detection that a relevant data processing behavior is in the second abnormal behavior list during data processing and / or the aforementioned first permission monitoring thread detects an unauthorized access situation. A data alert pop-up window appears on the content generation page corresponding to the aforementioned information processing platform to handle the data alert. The second abnormal behavior can be any abnormal data processing behavior. Each second abnormal behavior in the second abnormal behavior list is constructed based on historical experience and data augmentation. When a relevant second abnormal behavior is detected being executed within the corresponding container content, a data alert pop-up window will appear on the content generation page to alert to the existence of the abnormal behavior and to promptly handle the abnormal behavior operation. In practice, the first abnormal behavior list may include: abnormal behavior of acquiring unauthorized tools for data processing, abnormal behavior that causes significant semantic deviations, and abnormal behavior of processing data that has not been authorized for processing.

[0036] Fifth, within the container corresponding to the aforementioned copywriting generation skill module, generate the copy corresponding to the data processing results, which will then be used as the release copy. The release copy can be a preliminary generation of content based on the data processing results. The container corresponding to the copywriting generation skill module can be the container of the execution environment corresponding to the copywriting generation skill module. That is, the copywriting generation operation is performed within this container. During the process of generating the copy corresponding to the data processing results, the copywriting generation skill module supports calling the large model of the first intelligent agent to achieve accurate content generation.

[0037] Step 6: Generate the initial release content based on the above release text.

[0038] As an example, the aforementioned implementing entity can directly determine the release text as the initial release content.

[0039] In some optional implementations of certain embodiments, the aforementioned generation skill modules further include: a music configuration skill module, an image configuration skill module, and a multi-element fusion skill module. The music configuration skill module can be a skill module for configuring the music played corresponding to the published text. The image configuration skill module can be a skill module for configuring the image accompanying the published text. The multi-element fusion skill module can be a skill module that merges the published text, the played music, and the configuration. The music configuration skill module can be a skill module that matches background music with similar style and semantics based on the semantic content of the published text. Specifically, the configuration logic corresponding to the music configuration skill module is as follows: obtaining the semantic content of the published text, obtaining a music feature vector library, querying similar music feature vectors from the music feature vector library based on the semantic content of the text, obtaining similar music, and scoring similar music. The configuration logic corresponding to the image configuration skill module is similar to that of the image configuration skill module. The fusion logic corresponding to the multi-element fusion skill module is as follows: music playback cropping, image cropping, and multi-element content enhancement generation (cropped music, cropped images, and published text are input into the large model for content generation).

[0040] Optionally, the aforementioned implementing entity may generate initial release content based on the aforementioned release document, including the following steps: The first step, in response to the detection that a relevant copywriting generation behavior is in the third abnormal behavior list during the copywriting generation process and / or the aforementioned first permission monitoring thread detects an unauthorized access situation, a data alarm pop-up window appears on the content generation page corresponding to the aforementioned information processing platform to handle the data alarm. The third abnormal behavior can be any abnormal behavior related to copywriting generation. Each of the third abnormal behaviors in the third abnormal behavior list is constructed based on historical experience and data augmentation. For example, a third abnormal behavior could be the extensive use of existing publicly available literature. Another example is the copywriting patchwork behavior.

[0041] The second step involves generating the following third step for the aforementioned release text: Sub-step 1: Within the container corresponding to the aforementioned music configuration skill module, generate the music corresponding to the aforementioned release text, and use it as the release music. The container corresponding to the music configuration skill module can be a container that executes the configuration process corresponding to the music configuration skill module. The release music can be music that matches the release text. For example, the release music can be music generated specifically for the release text's key points.

[0042] Sub-step 2, in response to the detection that a relevant music generation behavior is in the fourth abnormal behavior list during the music generation process and / or the aforementioned first permission monitoring thread detects an unauthorized access situation, a data alarm pop-up window appears on the content generation page corresponding to the aforementioned information processing platform to handle the data alarm. The fourth abnormal behavior can be an abnormal behavior in music generation. Each fourth abnormal behavior in the fourth abnormal behavior list is constructed based on historical experience and data augmentation. For example, the fourth abnormal behavior can be the behavior of copying music demos.

[0043] Sub-step 3 involves generating the image corresponding to the aforementioned publishing text within the container corresponding to the image configuration skill module, and using this image as the publishing image. The container corresponding to the image configuration skill module can be a container that executes the configuration process corresponding to the image configuration skill module. The publishing image can be an image that matches the publishing text. For example, the publishing image can be a background image for the publishing text. In practice, the publishing image can be generated based on a large model.

[0044] Sub-step 4: In response to the detection that a relevant image generation behavior is in the fifth abnormal behavior list during the image generation process and / or the aforementioned first permission monitoring thread detects an unauthorized access situation, a data alarm pop-up window appears on the content generation page corresponding to the aforementioned information processing platform to handle the data alarm. The fifth abnormal behavior can be an abnormal behavior of image generation. Each fifth abnormal behavior in the fifth abnormal behavior list is constructed based on historical experience and data augmentation. For example, the fifth abnormal behavior can be the unauthorized direct use of a publicly available image, or the generation of a violation-related image.

[0045] The third step involves generating initial content within the container corresponding to the aforementioned multi-element fusion skill module, based on the published image, music, and text. The multi-element fusion skill module can be a skill module that fuses multiple elements of content. In practice, multi-element content fusion can involve fusing text, images, music, and other elements to generate complete initial content including multimodal content. In practice, the container corresponding to the multi-element fusion skill module can also be a container that performs the fusion of multiple elements of content. That is, within the container corresponding to the multi-element fusion skill module, the fusion of text, images, music, and other elements is performed (for example, by using a large model to fuse multiple elements of content) to obtain the initial content.

[0046] Fourth, in response to the detection that a related image generation behavior is in the sixth abnormal behavior list during the multi-element fusion process and / or the aforementioned first permission monitoring thread detects an unauthorized access situation, a data alarm pop-up window appears on the corresponding content generation page of the aforementioned information processing platform to handle the data alarm. The sixth abnormal behavior can be a deviation of the semantics of the generated initial published content from the content of each element. The sixth abnormal behavior can be an abnormal behavior during multi-element content fusion. Each sixth abnormal behavior in the sixth abnormal behavior list is constructed based on historical experience and data augmentation methods.

[0047] In addressing the technical problems mentioned in the background section, and considering the application scenario—urgent content generation and publishing—the following technical issues arise: For each skill generation module, a separate container is used to execute its corresponding skill steps. With a large number of skill generation modules, this results in the use of numerous containers for skill step execution, causing delays and failing to meet the demands of urgent content generation and publishing. Based on the specific requirements of this application scenario—rapid content generation and publishing—we have decided to adopt the following solution: In some optional implementations of certain embodiments, after step 1012, the steps further include: The first step, in response to the aforementioned requirement that the time interval between the information release time and the current time be no less than the target duration, is to obtain the release task information corresponding to the aforementioned information release instruction. This release task information can be the task information of the instruction task corresponding to the information release instruction. The task content of the task corresponding to the release task information is how to execute the instruction content corresponding to the information release instruction to release the corresponding content.

[0048] The second step is to filter from the historical task release case set at least one historical task release case whose task content similarity to the aforementioned task release information is higher than the target similarity. Here, historical task release cases can be various task release cases published at a historical time. Historical time can be a time prior to the current time. A task release case can be the content of a summary of the task execution process corresponding to a released task. Task content similarity can be the degree of similarity between task content. In practice, task content similarity can be a value between 0 and 1. That is, the content similarity between each task content corresponding to at least one historical task release case and the task content corresponding to the released task information is higher than the target similarity. In other words, the task content corresponding to each historical task release case in at least one historical task release case is highly similar to the task content corresponding to the released task information. The target similarity can be a threshold used to measure whether the content is highly similar. In practice, the target similarity can be set based on historical experience. For example, the target similarity can be 0.9.

[0049] The third step is to determine the container orchestration information for each historical release task case in at least one of the aforementioned historical release task cases. This container orchestration information includes the combination method of each generated skill module. Specifically, the container orchestration information can be the arrangement of the execution containers for each generated skill module when the historical release task case is executed. For example, the historical release task case might include generated skill modules 1, 2, 3, and 4. The corresponding container orchestration information could be that "generated skill module 1 and generated skill module 2" are executed in the same container, "generated skill module 3" is executed in the same container, and "generated skill module 4" is executed in the same container. The container orchestration information for each historical release task case can be the same or different.

[0050] Fourth, based on at least one obtained container orchestration information, a third agent is used to generate container orchestration information corresponding to each generated skill module in the aforementioned generated skill module skill information, which serves as the target container orchestration information. The target container orchestration information can represent the container orchestration status when each generated skill module is executed. That is, the target container orchestration information can characterize the container in which each generated skill module is executed.

[0051] As an example, firstly, the aforementioned execution entity can filter out a set of container orchestration information groups from at least one set of container orchestration information that corresponds to container orchestration information with the same orchestration method. Specifically, each container orchestration information group within the set shares the same container orchestration method. Then, the container orchestration information group with the largest number of corresponding container coding information groups is identified as the target container orchestration information group. Finally, the orchestration method corresponding to the target container orchestration information group is determined as the target container orchestration information.

[0052] Fifth, for the above target container orchestration information, perform the following fifth processing step: Sub-step 1: Based on the target container orchestration information, group the generated skill modules to obtain generated skill module sequence groups. Each generated skill module sequence has a corresponding execution order, and all generated skill modules within a sequence reside in the same container. The target container orchestration information characterizes how each generated skill module is allocated to a container for execution. Within the same container, each generated skill module in a sequence is executed sequentially. There is also a sequential execution order among the generated skill module sequences.

[0053] Sub-step 2: For each generation skill module sequence in the above-mentioned generation skill module sequence group, based on the mapping relationship between the generation skill module and the review skill module, determine the directed acyclic graph (DAG) corresponding to the above generation skill module sequence. In the DAG, the nodes are either generation skill modules or review skill modules, and the edges represent the execution relationships between the modules. The mapping relationship can be that after a generation skill module finishes execution, a certain review skill module must be executed to verify the accuracy of the output content corresponding to the generation skill module. For example, if the output content of a generation skill module is a processed dataset, then the corresponding review skill module is the skill module that performs data format verification on the processed dataset. The nodes in the DAG are either generation skill modules or review skill modules. The edges in the DAG can represent the execution order of the skill modules. In practice, the edges can be either the execution order between generation skill modules or the execution order between a generation skill module and a review skill module.

[0054] As an example, firstly, an initial directed acyclic graph (DAG) is generated based on the sequence of generated skill modules. This initial DAG represents the execution order of the modules corresponding to the generated skill modules. Then, based on the mapping relationship between the generated skill modules and the review skill modules, each review skill module corresponding to the sequence of generated skill modules is added to the initial DAG, resulting in the final DAG.

[0055] Sub-step 3 involves, based on the semantic parsing information described above, sequentially executing each directed acyclic graph (DAG) in the resulting DAG sequence by calling the first and second agents respectively, to obtain the published content. There is a one-to-one correspondence between the DAGs in the DAG sequence and the generated skill module sequences in the generated skill module sequence group. The specific implementation details are not elaborated further.

[0056] In some optional implementations of certain embodiments, the above-mentioned method of obtaining published content by sequentially executing each directed acyclic graph in the obtained directed acyclic graph sequence based on the semantic parsing information, by respectively calling the first agent and the second agent, includes the following steps: The first step involves inputting semantically parsed information into the first directed acyclic graph (DAG). This is achieved by calling the first and second agents respectively, following the graph order corresponding to the first DAG, to generate and review content, yielding partial review and output results. The partial review result can be the result of the second agent executing the corresponding review skill module. The partial output result can be the output of the first agent executing the generation skill module. The first DAG can be the first DAG in the sequence of DAGs, i.e., the first DAG to be executed.

[0057] The second step, in response to the aforementioned local audit result indicating the absence of audit issues, is to determine the next directed acyclic graph (DAG) in the first DAG as the target DAG. The target DAG can be the DAG to be executed by each node.

[0058] Third, for the target directed acyclic graph, perform the following sixth processing step: Sub-step 1: Obtain the previous directed acyclic graph (DAG) whose corresponding position in the target DAG is at the previous position. The previous position can be the previous position in the sequence of directed acyclic graphs.

[0059] Sub-step 2 determines the local output result corresponding to the previous directed acyclic graph. Details will not be elaborated further.

[0060] Sub-step 3 involves inputting the local output results into the target module's directed acyclic graph (DAG). By calling the first and second agents respectively, content generation and review are performed according to the graph order corresponding to the target DAG, resulting in the target local review results and target local output results.

[0061] Sub-step 4: In response to the above target local review result indicating that there is no review issue and the above target directed acyclic graph being the last directed acyclic graph in the module directed acyclic graph sequence, the target local output result is determined as the published content.

[0062] Fourth, in response to the above-mentioned target local audit result indicating that there is no audit problem and the above-mentioned target directed acyclic graph is not the last directed acyclic graph in the module directed acyclic graph sequence, the next directed acyclic graph corresponding to the target directed acyclic graph is determined as the target directed acyclic graph, and the above-mentioned sixth processing step is continued.

[0063] As another inventive aspect of this disclosure, the present disclosure first identifies at least one historical published task case that is highly similar to the task content of the published task information. Then, based on the container orchestration information corresponding to each of the at least one historical published task case, container arrangement information corresponding to each generated skill module can be accurately generated with reasonable and minimal resource consumption, resulting in a directed acyclic graph sequence. Finally, based on the order of each directed acyclic graph in the sequence, execution is performed sequentially, enabling rapid and efficient generation and publication of published content while ensuring access control and minimizing resource consumption.

[0064] In adopting technical solutions to address the technical problems mentioned in the background, the application scenario—requiring urgent content generation and publishing—presents the following technical challenge: when executing sequentially according to a directed acyclic graph, how to efficiently and accurately handle node-specific issues when review problems arise becomes the primary problem to solve. Considering the specific requirements of this application scenario—rapid content generation, rapid content publishing, and the existence of review issues—we have decided to adopt the following solution: In some optional implementations of certain embodiments, after determining the next directed acyclic graph in the first directed acyclic graph as the target directed acyclic graph in response to the local audit result indicating that no audit issue exists, the method further includes: In response to the aforementioned partial audit results indicating audit issues, the following seventh processing step is executed: Sub-step 1: Determine the number of adjustments corresponding to the first directed acyclic graph. The number of adjustments can be the number of times the graph content of the first directed acyclic graph is adjusted.

[0065] Sub-step 2, in response to the fact that the number of adjustments is less than the first count, locate the nodes corresponding to the aforementioned review issues. Here, the first count can be a value used to measure whether there are fewer adjustments to the directed acyclic graph. For example, the first count could be 3. Setting the first count here ensures that multiple adjustments to the directed acyclic graph are avoided, preventing longer output times. Each node can be a node that may have output problems. That is, after review by the subsequent review skills module, problems are found in the output content of each node.

[0066] Sub-step 3 involves determining the preceding node for each node in the first directed acyclic graph. The preceding nodes can be nodes generated from the input results of each node.

[0067] As an example, for each node in the series, identify the node that has an edge connection to the previous node and designate it as the preceding node.

[0068] Sub-step 4 involves deduplicating each of the preceding nodes to obtain a deduplicated node set.

[0069] Sub-step 5: Obtain the node output result set corresponding to the above deduplication node set. There is a one-to-one correspondence between the deduplication nodes in the deduplication node set and the node output results in the node output result set. The node output results can be the output results of the deduplication nodes performing content output.

[0070] Sub-step 6 involves using a third-party agent to adjust the nodes mentioned above, resulting in the adjusted nodes. Here, node adjustment can involve modifying the generation logic of the corresponding skill generation module for each node. There is a one-to-one correspondence between the nodes in the original nodes and the nodes in the adjusted nodes.

[0071] As an example, for each node within each node, firstly, an adjustment prompt is generated based on the review sub-question corresponding to the node, indicating an adjustment to the node's corresponding generation logic. Then, the adjustment prompt is input into the third agent to obtain the adjusted node.

[0072] Sub-step 7: Based on the adjusted nodes, update the first directed acyclic graph to obtain the first updated graph.

[0073] As an example, the aforementioned execution entity can replace each node in the first directed acyclic graph with the adjusted nodes to obtain the first updated graph.

[0074] Sub-step 8 involves inputting the output result sets of the above nodes into the corresponding nodes in the first update graph above, in order to continue content generation and content review.

[0075] In some optional implementations of certain embodiments, after locating the nodes corresponding to the audit issues in response to the number of adjustments being less than a predetermined number, the method further includes: The first step is to locate the relevant nodes for the aforementioned review issues in response to the number of adjustments reaching the first threshold.

[0076] The second step is to determine the target node set whose edge spacing is the target value. In practice, edge spacing can be the number of edges connecting nodes in a directed acyclic graph. For example, if node 1 and node 2 are connected by edge 1, and node 2 and node 3 are connected by edge 2, then the edges between node 1 and node 3 include edge 1 and edge 2, with a corresponding edge spacing of 2. Therefore, the edge spacing between node 1 and node 2 is 1. The target node set can be any set of nodes connected by edges with an edge spacing between them.

[0077] The third step is to perform deduplication on the above target node set to obtain a deduplicated target node set.

[0078] The fourth step is to obtain the target node output result set corresponding to the above deduplication target node set. There is a one-to-one correspondence between the deduplication target nodes in the deduplication target node set and the target node output results in the target node output result set.

[0079] The fifth step involves using a third-party intelligent agent to generate and adjust the target node set described above, resulting in the adjusted target node set. Details will not be elaborated further.

[0080] Step 6: Based on the adjusted target node set, update the first directed acyclic graph to obtain the second updated graph.

[0081] The seventh step involves inputting the output sets from the target nodes into the corresponding nodes in the second update graph to continue content generation and content review. Further details are omitted.

[0082] The above content is another inventive aspect of this disclosure. This disclosure adjusts the original directed acyclic graph (DAG) by setting a set number of adjustments for the DAG and continuously updating the local graph content in the DAG with node problems to obtain a first updated graph. This ensures the approval of the review process. Furthermore, after the first number of adjustments is reached, it indicates that there may be numerous problems in the local graph content. Therefore, it is necessary to expand the local graph content based on edge spacing to broaden the scope of graph structure adjustments and ensure approval.

[0083] Step 102: For the aforementioned initial published content, the second intelligent agent executes the first processing step: Step 1021: Obtain the review skills module information corresponding to the above-mentioned publishing platform.

[0084] In some embodiments, the aforementioned executing entity can obtain the review skill module information corresponding to the aforementioned publishing platform. The second intelligent agent can be an intelligent agent specifically responsible for reviewing the generated published content. In practice, the second intelligent agent has a corresponding large-scale model. The second intelligent agent can support processing multimodal content. The second intelligent agent not only performs basic sensitive word filtering but also identifies hidden risks through semantic understanding, contextual analysis, and multimodal cross-validation, making it particularly suitable for high-frequency publishing scenarios (e.g., social media operations, news media). The second intelligent agent can be an intelligent agent that reviews the generated published content in various aspects according to the platform requirements of the publishing platform. The large-scale model corresponding to the second intelligent agent can be a private review model. The second intelligent agent has a corresponding local knowledge base (i.e., access to an internal policy document library and a historical violation case library) and a corresponding security gateway (i.e., forcing all content to be reviewed before allowing calls to the publishing API). Specifically, different publishing platforms have different platform requirements and varying degrees of leniency in review. Therefore, for different publishing platforms, corresponding review skill module information can be set to achieve better publishing results on the corresponding publishing platform while quickly generating published content. Specific review aspects can include: explicit risk review and implicit risk review. Explicit risk review can include: automatically blocking explicitly defined violations such as political, violent, and pornographic content (e.g., sensitive word matching). Implicit risk review can include: identifying metaphorical violations through semantic reasoning. Review skill module information can be the module information corresponding to each review skill module. Review skill modules can be task skill modules used to review published content under the corresponding risk detection direction. That is, each review skill module has a corresponding risk detection direction. In practice, each review skill module can include: review skill modules corresponding to factual accuracy, compliance risk, value orientation, and technical defects. Factual accuracy can be demonstrated through data and case authenticity, time sensitivity, and key content comparison. Data and case authenticity can be: verifying whether cited data is properly attributed and whether cases actually exist (e.g., "A survey shows 80% user satisfaction" requires verification of the original report). Time sensitivity can be achieved by avoiding the use of outdated policy statements. Key content comparison can be achieved by forcibly calling authoritative source databases for comparison of content involving professional fields such as medicine and finance. Compliance risks can be identified through violations of advertising laws, copyright issues, and whether the rules are platform-specific. Value orientation can be reflected through implicit violations, cultural sensitivities, and essential capabilities. Implicit violations can include identifying "soft negative guidance." Cultural sensitivities can involve checking whether expressions related to ethnicity or religion comply with regulations (e.g., the accuracy of minority terminology). Essential capabilities can include performing contextual risk amplification tests on ambiguous expressions (e.g., whether changing the subject triggers a violation).Technical flaws can be identified through AI generation traces, homogenization risks, and format compatibility. AI generation traces can be detected by identifying overly neat sentence structures and a lack of emotional variation, thus preventing the platform from recognizing it as low-quality AI content. Homogenization risks can be mitigated by comparing historical content libraries and automatically marking content with a duplication rate exceeding 30% with traffic-limiting warnings. Format compatibility can be addressed by verifying whether special symbols (e.g., combinations of emojis) display abnormally on the publishing platform.

[0085] As an example, the aforementioned implementing entity can obtain the review skill module information corresponding to the aforementioned publishing platform by querying the mapping relationship table between the platform and the skill module.

[0086] Step 1022: Using the second two-layer sandbox architecture, the various review skill modules in the above review skill module information are called in sequence to review the content of the initial release and obtain the review results.

[0087] In some embodiments, the aforementioned executing entity can utilize a second two-layer sandbox architecture to sequentially invoke each review skill module in the aforementioned review skill module information to review the initially published content and obtain review results. The second two-layer sandbox architecture is a two-layer sandbox architecture set up for each review skill module. The review results may include: whether review issues exist and the specific review issues. Review issues may be defects found in the initially published content.

[0088] As an example, based on the second two-layer sandbox architecture and the initial published content, the second agent is instructed to execute the audit results sequentially according to the execution order of each audit skill module.

[0089] Step 103: In response to the above review results indicating the existence of review issues, the third-party intelligent agent bound to the above information processing platform is used to generate adjusted content for the above-mentioned initially published content.

[0090] In some embodiments, in response to the aforementioned review result indicating the existence of review issues, the executing entity utilizes a third intelligent agent bound to the aforementioned information processing platform to generate adjusted content for the initially published content. The third intelligent agent can be an intelligent agent bound to the information processing platform that supports various information processing functions. In practice, there is a one-to-one correspondence between various information processing functions and the various information processing tasks corresponding to the information processing platform. In practice, each information processing task can include: published content correction task, review process correction task, published content generation task, and review task. Specifically, the various task functions implemented by the third intelligent agent include the various task functions implemented by the first intelligent agent and the various task functions implemented by the second intelligent agent. The large model corresponding to the third intelligent agent is trained based on the training datasets corresponding to each information processing task. The model generation accuracy of the first intelligent agent is no higher than that of the model generation accuracy of the third intelligent agent. The model generation accuracy of the second intelligent agent is no higher than that of the model generation accuracy of the third intelligent agent. The first and second intelligent agents output for specific tasks, and their output speed is higher than that of the third intelligent agent. Furthermore, the update frequency of the first and second intelligent agents is higher than that of the third intelligent agent. The complexity of the large model corresponding to the third intelligent agent is higher than that of the large model corresponding to the first intelligent agent and the large model corresponding to the second intelligent agent. In practice, the large model corresponding to the third intelligent agent can be a combination of the large models corresponding to the first and second intelligent agents. Adjustments can be made to address issues in the initially released content to improve its quality. In practice, adjustments can be made to the released content by the first intelligent agent or to the review process by the second intelligent agent.

[0091] As an example, firstly, a prompt message is generated to predict the adjustments to the initially published content based on the review questions. Then, the prompt message is input into the larger model corresponding to the third agent to obtain the adjusted content for the initially published content.

[0092] In some optional implementations of certain embodiments, the aforementioned executing entity may utilize a third intelligent agent bound to the aforementioned information processing platform to generate adjusted content for the aforementioned initial published content, including the following steps: The first step is to use the aforementioned third agent to perform the second generation step: Sub-step 1: Obtain the review records corresponding to the above review results. These review records can be records of the initial published content being reviewed to obtain the review results. The review records may include: the execution records of the review skills corresponding to each review skill module.

[0093] As an example, a third-party intelligent agent can be used to generate a query statement for audit records to obtain the audit records corresponding to the audit results.

[0094] Sub-step 2 involves determining whether there are any process issues in the audit process corresponding to the aforementioned audit records. These process issues can be problems that arise during the audit process of each audit skill module. Each audit skill module has a specific execution order.

[0095] As an example, based on a larger model that is more accurate and comprehensive than that of the second agent corresponding to the third agent, it is determined whether there are process problems in the review process corresponding to the above review records.

[0096] Sub-step 3: In response to the existence of process issues, extract the subset of review skill modules corresponding to the aforementioned process issues from the review results. The review skill modules in this subset can be those that exhibited review logic issues during the initial review of the published content.

[0097] As an example, a subset of audit skill modules corresponding to the above process issues can be extracted from the audit results based on a larger model that is more accurate and comprehensive than that of the second agent, corresponding to the third agent.

[0098] Sub-step 4: Based on the above process issues, adjust the aforementioned subset of audit skill modules to obtain the adjusted subset of audit skill modules. There is a one-to-one correspondence between the adjusted audit skill modules in the adjusted subset and the audit skill modules in the original audit skill module subset.

[0099] As an example, based on the above process problem, the above audit skill module subset is adjusted using the corresponding large model of the third intelligent agent to obtain the adjusted audit skill module subset.

[0100] Sub-step 5 involves sending the first adjusted skill information and the first skill position corresponding to the adjusted subset of review skill modules to the second intelligent agent. The first adjusted skill information can be the summary of adjustments corresponding to the adjusted subset of review skill modules. That is, the first adjusted skill information can be the adjustment status corresponding to the adjusted subset of review skill modules. The first skill position can be the review execution position corresponding to each adjusted review skill module in the adjusted subset of review skill modules. In practice, each review skill module has a corresponding review execution position (i.e., representing the execution timing of the review skill module). Reviewing review skills executed before the first skill position does not present review problems. Reviewing review skills executed after the first skill position may present review problems. In practice... The second step involves executing a second processing step in response to the second agent obtaining the first adjustment skill information and the first skill location: Sub-step 1: Obtain the review snapshot corresponding to the initially published content. The review snapshot can be a detailed review of the initially published content. The review details corresponding to the review snapshot are more extensive than those corresponding to the review record. In practice, the review snapshot can be stored on the storage side corresponding to the second intelligent agent.

[0101] Sub-step 2: Obtain the intermediate review result corresponding to the first skill position from the aforementioned review snapshot. The intermediate review result can be the review result for the first skill position. Specifically, the intermediate review result is the sub-result that detects no review issues.

[0102] Sub-step 3: Based on the intermediate review results, using the second two-layer sandbox architecture, sequentially execute the adjusted review skill module subsets to obtain a re-review result. The re-review result can be based on the intermediate review results, using the second two-layer sandbox architecture to sequentially execute each adjusted review skill module in the adjusted review skill module subset, resulting in a re-review.

[0103] The third step is to send the results of the second review to the third intelligent agent, so that adjustments to the initial release content can be generated based on the results of the second review.

[0104] In some optional implementations of certain embodiments, the aforementioned executing entity may generate adjusted content for the initial release content based on the aforementioned re-review results, including the following steps: The first step is to respond to the above re-review result indicating that there is no review issue, generate adjustment content with an empty indicator, and determine the above initial release content as the release content.

[0105] The second step, in response to the above-mentioned re-audit results indicating the existence of audit issues, is to proceed with the third processing step: Sub-step 1: Based on the above review issues, locate at least one content generation skill module among the above-mentioned generation skill modules that has content generation issues.

[0106] As an example, firstly, semantic parsing is performed on the review questions to obtain various sub-questions. Then, for each sub-question, a group of generating skill modules that may generate sub-questions is identified. Finally, the obtained groups of generating skill modules are deduplicated to obtain at least one generating skill module.

[0107] Sub-step 2 involves determining the number of parameter adjustments for each of the at least one generated skill module. Each generated skill module has configurable parameters, which are automatically adjusted based on feedback. The number of parameter adjustments can be the total number of adjustments made to the parameters within the generated skill module. Specifically, during the content generation and publishing process, the number of parameter adjustments for each generated skill module is cumulative. For example, for the generated skill module "Hot Topic Capture Skill," the corresponding parameter adjustment could be to automatically reduce the weight of "Entertainment Hot Topics" from 0.3 to 0.1 and increase the weight of "Technology Hot Topics" from 0.5 to 0.7 if the click-through rate of "Entertainment Hot Topics" remains lower than that of "Technology Hot Topics." As another example, the copywriting generation skill has a "creativity_level" parameter (between 0 and 1, defaulting to 0.5). If the review results predict that copywriting with a creativity_level > 0.7 will have a higher interaction rate, this parameter can be automatically adjusted to 0.7.

[0108] As an example, the number of parameter adjustments for each generated skill module can be obtained from the temporary cache corresponding to the generated skill module information.

[0109] Sub-step 3, in response to the existence of a target generation skill module among the at least one generated skill module whose corresponding parameter adjustment count exceeds the target count, utilizes the aforementioned third agent to perform module semantic adjustment on the obtained at least one target generation skill module, resulting in at least one first adjusted module, and to adjust the parameters of the generation skill modules in the remaining generation skill module set, resulting in at least one second adjusted module. The target count can be used to measure whether multiple adjustments have yielded poor review results. In practice, the target count can be set based on historical experience. For example, the target count could be 5. By setting a target count, resource waste caused by multiple parameter adjustments can be avoided. The remaining generation skill module set can be the individual generation skill modules remaining after removing at least one target generation skill module from the at least one generation skill module. The second adjusted module can be a module whose parameters in the remaining generation skill modules have been adjusted.

[0110] Here, a third agent corresponding to the large model can be used to perform module semantic adjustment on at least one target-generated skill module to obtain at least one first-adjusted module. In addition, based on the review sub-results corresponding to each remaining generated skill module in the set of remaining generated skill modules, the parameters in the remaining generated skill modules are dynamically adjusted using the third agent to obtain a second-adjusted module.

[0111] Sub-step 4 involves sending the module adjustment information and the second skill position for at least one first adjusted module and at least one second adjusted module to the first agent. The module adjustment information can represent the overall adjustment content of the modules corresponding to the at least one first adjusted module and the at least one second adjusted module. The module adjustment information can also represent the overall adjustment content corresponding to at least one generated skill module. The second skill position can be the skill module position corresponding to at least one first adjusted module and at least one second adjusted module.

[0112] Third, in response to the first intelligent agent obtaining the second adjustment skill information and the second skill location, the fourth processing step is executed.

[0113] Sub-step 1: Obtain the snapshot corresponding to the initial published content. This snapshot can be a snapshot of the detailed process of generating the initial published content. In practice, the snapshot can be stored on the storage side corresponding to the first intelligent agent.

[0114] Sub-step 2 involves obtaining the intermediate generated content corresponding to the second skill position from the aforementioned generated snapshot. This intermediate generated content can be the publishing-related content generated in the snapshot generated at the second skill position.

[0115] As an example, first, a content search statement for the second skill position is generated. Then, the content search statement is executed to retrieve intermediate generated content from the aforementioned generated snapshot.

[0116] Fourth, based on the intermediate generated content, the adjusted skill module sub-sequences are executed sequentially using the first two-layer sandbox architecture to obtain regenerated content, which is then used as the adjusted content.

[0117] As an example, the aforementioned execution entity can, under the container constraints corresponding to the first two-layer sandbox architecture, substitute the intermediate generated content into the sub-sequence of the generated skill module, and execute the generated skill module sequentially according to the sequence order to obtain the regenerated content as the adjustment content.

[0118] Step 104: Using the first intelligent agent and / or the second intelligent agent described above, execute the above-mentioned adjustments to obtain the published content.

[0119] In some embodiments, the executing entity may utilize the first intelligent agent and / or the second intelligent agent to execute the aforementioned adjustments and obtain the published content. The published content may be content that has been confirmed and supported by both the first and second intelligent agents before publication.

[0120] As an example, in response to the adjustment content being for a first agent, the first agent is instructed to execute the adjustment, resulting in the initial release of the adjustment content. Then, a second agent reviews the initial release of the adjustment content. Upon passing the review, the initial release of the adjustment content is confirmed as published. In response to the adjustment content being for a second agent, the second agent is instructed to execute the adjustment, resulting in reviewed adjustment content. In response to the reviewed adjustment content indicating that the review is correct, the initial release content is confirmed as published content.

[0121] Step 105: In response to receiving feedback from the review terminal regarding the above-mentioned published content and confirming that the feedback from the review terminal is correct, the above-mentioned published content is published to the above-mentioned publishing platform using a temporary token mechanism.

[0122] In some embodiments, in response to receiving feedback from the review terminal regarding the published content and confirming that the feedback is correct, the executing entity may use a temporary token mechanism to publish the published content to the publishing platform. The review terminal feedback result can be the result of the review terminal reviewing the published content. The review terminal can be the backend review terminal corresponding to the information processing platform. In practice, the review terminal can be a manual review terminal. The review terminal feedback result can be one of the following: a result indicating no error, or a result indicating an error. In the temporary token mechanism, long-term credentials refer to identity authentication credentials with long-term validity (e.g., permanent API keys, fixed tokens), while temporary tokens are short-term access credentials with time limits and dedicated to a single task. In scenarios such as MCN operation platforms, long-term credentials pose high security risks and have weak access control. The temporary token mechanism, through time-limited expiration and a least-permission design, significantly reduces the risks of account theft and data leakage.

[0123] In some optional implementations of certain embodiments, after step 105, the steps further include: The first step is to obtain both explicit and implicit feedback information regarding the published content. Explicit feedback information can be content that provides clear feedback on the published content. For example, explicit feedback information could be likes or positive comments on the published content. Implicit feedback information can be feature content expressed by the implicit characteristics of the published content. In practice, implicit feedback information may include: dwell time.

[0124] The second step involves generating content summary information based on the aforementioned explicit and implicit feedback information, using the third intelligent agent. This content summary information can be a compilation of the content quality and feedback status of the published content.

[0125] As an example, the aforementioned executing entity can generate summary prompts based on explicit and implicit feedback information to summarize content and assess content quality. These summary prompts are then input into the corresponding model of the third agent to obtain the content summary information.

[0126] The third step is to summarize the information based on the above content, perform semantic adjustments on the generated skill module information and the reviewed skill module information to obtain module adjustment information, which will be cached in the model database.

[0127] As an example, firstly, a keyword set is extracted from the content summary information. Then, for each keyword, the corresponding skill module information group is determined. Finally, the obtained skill module information groups are deduplicated to obtain a skill module information set. From the skill module information set, target skill module information groups with a corresponding keyword count higher than the target count are selected. Using a third-party intelligent agent, based on the content summary information, the semantic adjustment of each target skill module information in the target skill module information group is performed to obtain module adjustment information.

[0128] The fourth step involves retraining the large model corresponding to the third agent based on the cached model datasets, resulting in an updated third agent. The specific training method will not be elaborated further.

[0129] Fifth, based on the updated third agent, update the first and second agents, and implement the module adjustment information.

[0130] As an example, the aforementioned execution entity can update the first and second intelligent agents and deploy the module adjustment information based on transfer learning and dataset generation.

[0131] The above-described embodiments of this disclosure have the following beneficial effects: Through the information publishing method based on multi-agent collaboration in some embodiments of this disclosure, multiple agents can efficiently and accurately publish the generated content to the relevant platform after inputting the corresponding information publishing instruction on the information processing platform. Specifically, the inefficiency of related content publishing stems from the following: complex and precise prompts need to be set based on a basic dataset, resulting in high cost and time consumption for prompt generation. Furthermore, the content output by large models contains a significant amount of sensitive content (e.g., hallucination content and copyright-infringing content), leading to substantial defects in the published content. Multiple refinements of prompts are required, which is time-consuming. Finally, the generation, uploading, and review of published content all require actual execution of related page processing and review operations, making the overall content publishing process cumbersome and time-consuming, unsuitable for large-scale content publishing. Based on this, some embodiments of the information publishing method based on multi-agent collaboration disclosed herein firstly, in response to selecting generation skill module information on the information processing platform and obtaining the input information publishing instruction, the first agent performs the following first generation step: Step 1: Semantic content parsing of the aforementioned information publishing instruction yields semantic parsing information, which includes publishing platform and publishing content requirement information. Here, the information processing platform is set up so that skill module information and information publishing instructions can be selected on the information processing platform to accurately and efficiently generate the required publishing content based on the information publishing requirements and the content generation task module. Based on the first agent, by utilizing the generation skill module information, various logical modules for generating publishing content are automatically invoked, enabling efficient and streamlined generation of publishing content. By performing semantic content parsing on the information publishing instruction, the content requirements and publishing platform information used for publishing content generation can be obtained. Step 2: In response to the aforementioned publishing content requirement information indicating that the time interval between the publishing time and the current time is less than the target duration, based on the aforementioned semantic parsing information, each generation skill module in the aforementioned generation skill module information is sequentially invoked using a first two-layer sandbox architecture to generate the initial publishing content under the aforementioned publishing platform. Here, when the required information representation for the published content is shorter than the target duration, a comprehensive processing approach can be adopted: first, an initial published content is generated by a first intelligent agent, and then a second intelligent agent reviews the initial published content. This significantly reduces the efficiency of both generation and review. Furthermore, utilizing the first two-layer sandbox architecture avoids permission violations during the invocation of various generation skill modules, preventing financial losses. By sequentially executing each generation skill module, the initial published content for the publishing platform can be generated automatically and accurately.Then, for the aforementioned initial published content, the second intelligent agent executes the first processing step: First, it obtains the review skill module information corresponding to the aforementioned publishing platform, so as to automatically and accurately review the initial published content based on each review skill module, thereby ensuring the quality of the initial published content. Second, using the second two-layer sandbox architecture, it sequentially calls each review skill module in the aforementioned review skill module information to review the content of the aforementioned initial published content and obtain the review result. Here, the use of the second two-layer sandbox architecture can avoid the problem of unauthorized access and avoid loss of property. Moreover, using each review skill module can accurately and efficiently perform semantic review of the content of the initial published content. Next, in response to the aforementioned review result indicating the existence of review issues, the third intelligent agent bound to the aforementioned information processing platform generates adjusted content for the aforementioned initial published content. Here, the third intelligent agent can make timely adjustments to the content with review issues to assist in the accurate generation of high-quality published content subsequently. Then, using the aforementioned first intelligent agent and / or the aforementioned second intelligent agent, the aforementioned adjusted content is executed to obtain high-quality published content. Finally, in response to receiving feedback from the review terminal regarding the aforementioned published content, and confirming that the feedback is accurate, the published content is then published to the aforementioned publishing platform using a temporary token mechanism. This temporary token mechanism, through its time-limited expiration and least privilege design, significantly reduces the risks of account theft and data leakage. In summary, through various intelligent agents with different functions, after inputting corresponding information publishing commands on the information processing platform, the generated content can be automatically and efficiently published to the relevant platforms.

[0132] Further reference Figure 2 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of an information publishing device based on multi-agent collaboration. These device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, this information publishing device based on multi-agent collaboration can be specifically applied to various electronic devices.

[0133] like Figure 2As shown, an information publishing device 200 based on multi-agent collaboration includes: a first execution unit 201, a second execution unit 202, a generation unit 203, a third execution unit 204, and a publishing unit 205. The first execution unit 201 is configured to, in response to selecting generation skill module information on an information processing platform and obtaining an input information publishing instruction, utilize a first agent to perform the following first generation step: semantic content parsing of the information publishing instruction to obtain semantic parsing information, wherein the semantic parsing information includes: publishing platform and publishing content requirement information; in response to the publishing content requirement information indicating that the time interval between the publishing time and the current time is less than the target duration, based on the semantic parsing information, using a first two-layer sandbox architecture, sequentially calling each generation skill module in the generation skill module information to generate initial publishing content under the publishing platform; the second execution unit 202 is configured to, for the initial publishing content, utilize a second agent to perform a first processing step: obtaining... The following steps are taken: 1) Obtain the review skill module information corresponding to the aforementioned publishing platform; 2) Utilize the second two-layer sandbox architecture to sequentially call each review skill module in the aforementioned review skill module information to review the initial published content and obtain the review result; 3) Generate unit 203 is configured to, in response to the aforementioned review result indicating a review problem, use the third intelligent agent bound to the aforementioned information processing platform to generate adjusted content for the initial published content; 4) Third execution unit 204 is configured to, using the aforementioned first intelligent agent and / or the aforementioned second intelligent agent, execute the adjusted content to obtain the published content; 5) Publishing unit 205 is configured to, in response to receiving the review terminal feedback result for the published content and the review terminal feedback result indicating no error, use a temporary token mechanism to publish the published content to the aforementioned publishing platform.

[0134] It is understandable that the units and references described in the multi-agent collaborative information publishing device 200 are... Figure 1 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the information publishing device 200 based on multi-agent collaboration and the units contained therein, and will not be repeated here.

[0135] The following is for reference. Figure 3 It shows a schematic diagram of the structure of an electronic device (e.g., an electronic device) 300 suitable for implementing some embodiments of the present disclosure. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0136] like Figure 3As shown, the electronic device 300 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0137] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 3 Each box shown can represent a device or multiple devices as needed.

[0138] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 309, or installed from storage device 308, or installed from ROM 302. When the computer program is executed by processing device 301, it performs the functions defined in the methods of some embodiments of this disclosure.

[0139] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0140] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0141] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: in response to selecting generation skill module information on the information processing platform and obtaining an input information publishing instruction, utilize a first intelligent agent to perform the following first generation step: perform semantic content parsing on the aforementioned information publishing instruction to obtain semantic parsing information, wherein the semantic parsing information includes: publishing platform and publishing content requirement information; in response to the aforementioned publishing content requirement information indicating that the time interval between the publishing time and the current time is less than the target duration, based on the aforementioned semantic parsing information, utilize a first two-layer sandbox architecture to sequentially call each generation skill module in the aforementioned generation skill module information to generate initial publishing content under the aforementioned publishing platform; Regarding the aforementioned initial published content, the second intelligent agent performs the first processing step: obtaining the review skill module information corresponding to the aforementioned publishing platform; using the second two-layer sandbox architecture, sequentially calling each review skill module in the aforementioned review skill module information to review the aforementioned initial published content and obtain the review result; in response to the aforementioned review result indicating the existence of review issues, using the third intelligent agent bound to the aforementioned information processing platform, generating adjusted content for the aforementioned initial published content; using the aforementioned first intelligent agent and / or the aforementioned second intelligent agent, executing the aforementioned adjusted content to obtain the published content; in response to receiving the review terminal feedback result for the aforementioned published content and the review terminal feedback result indicating no errors, using a temporary token mechanism, publishing the aforementioned published content to the aforementioned publishing platform.

[0142] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0143] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0144] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a first execution unit, a second execution unit, a generation unit, a third execution unit, and a publishing unit. The names of these units do not necessarily limit the specific unit; for example, the third execution unit may also be described as "a unit that uses the aforementioned first intelligent agent and / or the aforementioned second intelligent agent to execute the aforementioned adjustment content and obtain the published content."

[0145] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0146] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for information dissemination based on multi-agent collaboration, comprising: In response to selecting and generating skill module information on the information processing platform and receiving the input information publishing instruction, the first intelligent agent executes the following first generation step: The information publishing instruction is semantically parsed to obtain semantic parsing information, wherein the semantic parsing information includes: publishing platform and publishing content requirements information; In response to the requirement that the time interval between the release time and the current time is less than the target duration, based on the semantic parsing information, the first two-layer sandbox architecture is used to sequentially call each generation skill module in the generation skill module information to generate the initial release content under the release platform. For the initially published content, the second intelligent agent executes the first processing step: Obtain the review skills module information corresponding to the publishing platform; Using the second two-layer sandbox architecture, each review skill module in the review skill module information is called sequentially to review the initially published content and obtain the review result; In response to the audit result indicating the existence of audit issues, the third-party intelligent agent bound to the information processing platform is used to generate adjusted content for the initially published content; The adjustment content is executed using the first intelligent agent and / or the second intelligent agent to obtain the published content; Upon receiving feedback from the review terminal regarding the published content, and confirming that the feedback is accurate, the published content is published to the publishing platform using a temporary token mechanism.

2. The method according to claim 1, wherein, The step of using a third-party intelligent agent bound to the information processing platform to generate adjusted content for the initially published content includes: Using the third intelligent agent, perform the second generation step: Retrieve the audit record corresponding to the audit result; Determine if there are any process issues in the review process corresponding to the review record; In response to the existence of process issues, a subset of audit skill modules corresponding to the process issues is extracted from the audit results; Based on the process issues, the subset of audit skills modules is adjusted to obtain the adjusted subset of audit skills modules. Send the first adjusted skill information and the first skill location corresponding to the adjusted audit skill module subset to the second intelligent agent; In response to the second agent obtaining the first adjusted skill information and the first skill position, the second processing step is executed: Obtain the review snapshot corresponding to the initially published content; Obtain the intermediate review result corresponding to the first skill position from the review snapshot; Based on the intermediate review results, the adjusted review skill module subsets are executed sequentially using the second two-layer sandbox architecture to obtain a second review result; The re-review result is sent to the third intelligent agent to generate adjusted content for the initially published content based on the re-review result.

3. The method according to claim 2, wherein, Based on the results of the second review, the generation of adjusted content for the initially published content includes: In response to the re-review result indicating that there is no review issue, an adjustment content with an empty representation is generated, and the initial release content is determined as the release content; In response to the re-audit result indicating an audit issue, a third processing step is executed: Based on the review issues, identify at least one content generation skill module among the various generation skill modules that has content generation issues; Determine the number of parameter adjustments for each of the at least one skill generation module; In response to the existence of a target generated skill module in the at least one generated skill module where the number of corresponding parameter adjustments is greater than the target number, the third agent is used to perform module semantic adjustment on the obtained at least one target generated skill module to obtain at least one first adjusted module, and to adjust the parameters of the generated skill modules in the remaining generated skill module set to obtain at least one second adjusted module. The module adjustment information and second skill position for the at least one first adjusted module and the at least one second adjusted module are sent to the first intelligent agent; In response to the first agent obtaining the second adjusted skill information and the second skill position, the fourth processing step is executed: Obtain a snapshot corresponding to the initially published content; Obtain the intermediate generated content corresponding to the position of the second skill from the generated snapshot; Based on the intermediate generated content, the adjusted generation skill module sub-sequence is executed sequentially using the first two-layer sandbox architecture to obtain regenerated content, which serves as the adjusted content.

4. The method according to claim 1, wherein, The method further includes: Obtain explicit and implicit feedback information regarding the published content; Based on the explicit feedback information and the implicit feedback information, the third intelligent agent generates content summary information. Based on the summarized information, the semantic adjustment of the generated skill module information and the reviewed skill module information is performed to obtain module adjustment information, which is then cached in the model database. Based on the cached model datasets, the large model corresponding to the third agent is retrained to obtain the updated third agent. Based on the updated third agent, the first agent and the second agent are updated, and the module adjustment information is uploaded.

5. The method according to claim 1, wherein, After performing semantic content parsing on the information publishing instruction to obtain semantic parsing information, the method further includes: Using the third intelligent agent, a first two-layer sandbox structure for generating skill module information and a second two-layer sandbox structure for reviewing skill module information are constructed. A first permission monitoring thread is configured to run when the first intelligent agent executes the generation of skill module information, and a second permission monitoring thread is configured to run when the second intelligent agent executes the review of skill module information; and The step of generating initial content for the publishing platform by sequentially calling each generation skill module in the generation skill module information based on the semantic parsing information and utilizing the first two-layer sandbox architecture includes: Based on the semantic parsing information, using the first two-layer sandbox architecture and the first permission monitoring thread, each generation skill module in the generation skill module information is called in sequence to generate the initial release content under the release platform. In this case, if the first permission monitoring thread detects that the first agent has exceeded its permission when executing the corresponding generation skill module, it directly triggers the first agent to pause processing until it restarts processing on the corresponding permission recovery page of the information processing platform. Based on the initial published content, the second two-layer sandbox architecture and the second permission monitoring thread are used to sequentially call each of the review skill modules in the review skill module information to review the initial published content and obtain the review result. In this case, if the second permission monitoring thread detects that the second agent has exceeded its authority when executing the corresponding review skill module, it directly triggers the second agent to pause processing until it is restarted on the corresponding permission recovery page of the information processing platform.

6. The method according to claim 5, wherein, The various skill generation modules include: a data acquisition skill module, a data processing skill module, and a copywriting generation skill module; and The step of generating initial content for the publishing platform by sequentially calling each generation skill module in the generation skill module information based on the semantic parsing information and utilizing the first two-layer sandbox architecture includes: Within the container corresponding to the data acquisition skill module, basic publishing data is acquired based on the publishing content requirements information; In response to the detection that the relevant data acquisition behavior is in the first abnormal behavior list during the acquisition process and / or the first permission monitoring thread detects that there is an unauthorized permission situation, a data alarm pop-up window will pop up on the content generation page corresponding to the information processing platform to handle the data alarm. Within the container corresponding to the data processing skill module, the published basic data is processed to obtain the data processing result; In response to the detection that the relevant data processing behavior is in the second abnormal behavior list during the data processing process and / or the first permission monitoring thread detects that there is an unauthorized permission situation, a data alarm pop-up window will pop up on the content generation page corresponding to the information processing platform to process the data alarm. Within the container corresponding to the copywriting generation skill module, the copywriting corresponding to the data processing result is generated and used as the publishing copywriting. Generate initial release content based on the aforementioned release text.

7. The method according to claim 6, wherein, The various generation skill modules also include: a music configuration skill module, an image configuration skill module, and a multi-element fusion skill module; and The step of generating initial release content based on the release text includes: In response to the detection that the relevant copy generation behavior is in the third abnormal behavior list during the copy generation process and / or the first permission monitoring thread detects that there is an overreach of permissions, a data alarm pop-up window will be displayed on the content generation page corresponding to the information processing platform to handle the data alarm. For the aforementioned release text, perform the following third generation step: Within the container corresponding to the music configuration skill module, generate the music corresponding to the release text as the release music; In response to the detection that the relevant music generation behavior is in the fourth abnormal behavior list during the music generation process and / or the first permission monitoring thread detects that there is an overreach of permissions, a data alarm pop-up window will be displayed on the content generation page corresponding to the information processing platform to process the data alarm. Within the container corresponding to the image configuration skill module, an image corresponding to the publishing text is generated and used as the publishing image; In response to the detection that the relevant image generation behavior is in the fifth abnormal behavior list during the image generation process and / or the first permission monitoring thread detects that there is an unauthorized permission situation, a data alarm pop-up window will be displayed on the content generation page corresponding to the information processing platform to process the data alarm. Within the container corresponding to the multi-element fusion skill module, initial publishing content is generated based on the publishing image, the publishing music, and the publishing text. In response to the detection that the relevant image generation behavior is in the sixth abnormal behavior list during the multi-element fusion process and / or the first permission monitoring thread detects that there is an unauthorized permission situation, a data alarm pop-up window will appear on the content generation page corresponding to the information processing platform to process the data alarm.

8. An information publishing device based on multi-agent collaboration, comprising: The first execution unit is configured to, in response to selecting skill module information on the information processing platform and receiving an input information publishing instruction, utilize a first intelligent agent to execute the following first generation step: perform semantic content parsing on the information publishing instruction to obtain semantic parsing information, wherein the semantic parsing information includes: publishing platform and publishing content requirement information; in response to the publishing content requirement information indicating that the time interval between the publishing time and the current time is less than the target duration, based on the semantic parsing information, utilize a first two-layer sandbox architecture to sequentially call each of the skill modules in the skill module information to generate the initial publishing content under the publishing platform; The second execution unit is configured to, for the initial published content, use a second intelligent agent to execute the first processing step: obtain the review skill module information corresponding to the publishing platform; and use a second two-layer sandbox architecture to sequentially call each review skill module in the review skill module information to review the initial published content and obtain the review result. The generation unit is configured to, in response to the review result indicating the existence of review issues, use a third-party intelligent agent bound to the information processing platform to generate adjusted content for the initially published content. The third execution unit is configured to use the first intelligent agent and / or the second intelligent agent to execute the adjustment content and obtain the published content. The publishing unit is configured to, in response to receiving feedback from the review terminal regarding the published content and confirming that the feedback from the review terminal is correct, publish the published content to the publishing platform using a temporary token mechanism.

9. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable medium having a computer program stored thereon, wherein, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.