A method and system for controlled distribution and feedback optimization of multimedia content

Abstract

The application discloses a multimedia content controlled publishing and feedback optimization method and system, which comprises the following steps: analyzing product source data to obtain structured product data with fact source identification, generating platform content according to platform constraints of a target platform; performing compliance verification and publishing executability verification on the platform content, returning correctable content for correction and blocking terminated content; binding the identification and version of the verified content with the target platform to generate a publishing task, and completing page filling, media uploading, submission and result confirmation by a platform execution module; collecting effect feedback of the published content and associating the corresponding content version, and updating subsequent generated control data without modifying product facts and prohibited items. The application can reduce content writing errors, format inadaptation, wrong versions and repeated publishing, and improve the automation degree and execution stability of cross-platform publishing.

Description

Technical Field

[0001] This invention relates to the field of computer data processing and network automation technology, specifically to a method and system for controlled publishing and feedback optimization of multimedia content. Background Technology

[0002] Currently, when enterprises promote and operate new media and multimedia platforms, they often need to convert the same product information into content with different title structures, body lengths, media proportions, hashtags, and publication times. Existing methods often rely on operations staff to extract information from each platform, rewrite content, upload media, and check the publication results. This leads to the repetitive processing of the same facts, difficulty in tracking task handover and version changes, and is inefficient, costly, and time-consuming.

[0003] With the application of large language models and multimodal generative models, some solutions can generate text, images, or videos based on prompts, or simulate manual page filling using automated tools. However, many problems still exist: First, the content generated by the model may be inconsistent with the product source materials, or violate platform content rules, legal regulations, and brand guidelines; second, the length, media size, required fields, or file format of the generated results may not be directly accepted by the target platform, causing the automation process to fail during the upload or submission stage; third, when the platform responds with a timeout or the page structure changes, simple retries may result in the same version being published repeatedly, and the exposure, click, and conversion data after publication are not accurately correlated with the original content version and cannot be used for subsequent generation.

[0004] Therefore, there is a need for a technical solution that checks content risks and release feasibility before submission, prevents errors and duplicate submissions during the execution phase, and uses release feedback to improve subsequent content generation, so as to improve the degree of automation while keeping the release process controllable, traceable and recoverable. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for controlled publishing and feedback optimization of multimedia content, in order to solve the problems of fragmented cross-platform content generation and publishing processes, difficulty in unifying and verifying content facts and rules, automatic publishing anomalies that easily lead to misprints or duplicate publishing, and the inability of effect feedback to recycle back to the content generation process.

[0006] To achieve the above objectives, the present invention provides a method for controlled publishing and feedback optimization of multimedia content, comprising the following steps: S1. Parse the product source data to obtain structured product data; S2. For at least one target platform, based on the structured product data and the platform constraint data corresponding to the target platform, generate platform content data through a multimodal content generation model; S3. Perform content compliance verification on the platform content data based on the content constraint data, and perform publication executability verification on the platform content data based on the platform constraint data. Based on the results of the two verifications, determine the platform content data as publishable content, content to be corrected, or terminated content; wherein, the content to be corrected is returned to the multimodal content generation model for correction, and the terminated content is blocked. S4. Bind the content identifier and content version in the publishable content to the target platform to generate a publishing task; S5. The platform execution module executes the publishing task and determines the publishing result based on the publishing status information returned by the target platform; S6. Collect the effect feedback data corresponding to the successfully published content, associate the effect feedback data with the publishable content according to the content identifier to obtain the effect evaluation result, and update the generation control data of subsequent platform content data according to the effect evaluation result under the condition of satisfying the content constraint data and the platform constraint data.

[0007] To optimize the above technical solution, the specific measures also include: Further, in step S1, the product source data includes at least one of local document data, online knowledge base data, and business system interface data; the structured product data includes at least two of product name, specifications, functional information, applicable scenarios, and brand information, and each field is associated with a fact source identifier, which is used to verify the consistency between factual statements and product source data during content compliance verification.

[0008] Furthermore, the platform constraint data mentioned in step S2 includes text length constraints, content structure constraints, media specification constraints, page element constraints, and publication completion judgment conditions; when generating the platform content data, at least one of the title, body text, topic tags, images, and videos is organized according to the content structure constraints, and corresponding media files are generated and converted according to the media specification constraints.

[0009] The media specification constraints include at least one of image size, image aspect ratio, media file size, video duration, video encoding format, and media container format. The page element constraints include required page elements in the target page, page element positioning information, and page element operability status.

[0010] Furthermore, the method is implemented through a product analysis intelligent agent, a content generation intelligent agent, a compliance review intelligent agent, and a task scheduling intelligent agent; the product analysis intelligent agent is used to assist in generating the structured product data, the content generation intelligent agent is used to call a multimodal content generation model to generate the platform content data, the compliance review intelligent agent is used to perform content compliance verification, and the task scheduling intelligent agent calls the platform execution module to execute the publishing task.

[0011] Further, in step S3, the content constraint data includes target platform rules, legal compliance rules, and brand specifications; the content compliance verification outputs three states: pass, correctable failure, and uncorrectable failure; the correctable failure verification items, their corresponding text positions, and modification instructions are written into the correction context, and the correction context and the content to be corrected are input into the multimodal content generation model; if the verification still fails after the number of corrections reaches a threshold, the corresponding content is determined as terminated content.

[0012] The content to be corrected and the corresponding correction context are input into the multimodal content generation model to generate corrected platform content data with incremental content versions, and step S3 is re-executed on the corrected platform content data; the terminated content is blocked so that it does not enter the publishing task generation step.

[0013] Furthermore, each platform's content data is assigned a content identifier and a corrected content version, and a task identifier is generated based on the content identifier, content version, target platform, and target account. Before resubmitting, the publishing record is queried based on the task identifier, and the resubmission operation is only performed if there is no successful publishing record.

[0014] Furthermore, the platform execution module connects to different target platforms. The platform execution module automatically loads mutually isolated account sessions that are still logged in, locates page elements of the target platform, performs text filling, media uploading and submission operations, and generates publishing status information indicating successful publishing when at least two of the following indicate successful publishing: publishing completion flag, publishing result address and remote publishing record identifier; it also generates a screenshot of the publishing result and saves it in association with the publishing status information.

[0015] Further, in step S6, the effect feedback data includes exposure, clicks, and conversions; based on the statistics of each target platform, the effect feedback data is normalized to obtain a content effect score, and at least one of the content structure parameters, media combination parameters, content length parameters, and release time parameters in the generation control data is adjusted according to the content effect score. At the same time, the fact fields in the structured product data and the prohibited items in the content constraint data are set as non-writable fields of the feedback optimization module, so that the feedback optimization module only updates the adjustable parameters in the generation control data.

[0016] This invention also provides a multimedia content controlled publishing and feedback optimization system, comprising: The data parsing module is equipped with a product analysis intelligent agent, which is used to parse the product source data to obtain structured product data; The content generation module is equipped with a content generation intelligent agent, which generates platform content data based on the structured product data and the platform constraint data corresponding to the target platform through a multimodal content generation model. The dual verification module is equipped with a compliance review intelligent agent and a release executability verifier. The compliance review intelligent agent is used to perform content compliance verification, and the release executability verifier is used to perform release executability verification. Based on the results of the two verifications, the module determines whether the content is publishable, content to be corrected, or content to be terminated. The task control module is equipped with a task scheduling agent to bind the content identifier and content version of the publishable content to the target platform to generate a publishing task; The platform execution module is used to execute the publishing task and determine the publishing result based on the publishing status information of the target platform; The feedback optimization module is used to associate the effect feedback data of successfully published content with the content identifier to obtain the effect evaluation result, and update the subsequent generation control data under the condition of satisfying the content constraint data and the platform constraint data.

[0017] Furthermore, the platform execution module is used to detect execution anomalies during the execution of the publishing task and output anomaly information to the task control module; the task control module is used to determine whether the execution anomaly is recoverable or unrecoverable based on the anomaly information; for the recoverable anomaly, after confirming that there is no successful publishing record based on the task identifier, the platform execution module is triggered to re-execute the corresponding publishing task according to a preset backoff interval; for the unrecoverable anomaly, the corresponding publishing task is terminated, and the task identifier, target platform, anomaly type, page status, screenshot address, and occurrence time are recorded.

[0018] The beneficial effects of this invention are as follows: First, product information from different sources is organized into structured data that allows for traceability of factual origins. When generating content, specific evidence can be verified, reducing instances of incorrect product parameters or unsubstantiated expansions. Before publication, content is simultaneously checked for compliance and platform format requirements. Issues that can be modified are automatically corrected, while content that cannot be published is directly blocked. During publication, the current task is confirmed through content version and task identifier. Before retrying, it checks whether publication has already been successful, thereby reducing errors and duplicate publications. Different platforms complete content entry, media upload, and result confirmation through their respective platform execution modules. Post-publication exposure, clicks, and conversion data are then associated with the original content version for adjusting subsequent content structure, media mix, and publication time. This reduces repetitive manual checks and operations, improving the stability and automation of cross-platform publishing. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the controlled publishing and feedback optimization method for multimedia content according to the present invention. Figure 2 This is a schematic diagram illustrating the dual verification and versioning correction in an embodiment of the present invention; Figure 3 This is a schematic diagram of the closed loop of versioned release, idempotent confirmation, and feedback optimization in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the multimedia content controlled publishing and feedback optimization system according to an embodiment of the present invention. Detailed Implementation

[0020] The invention will now be described in further detail with reference to the accompanying drawings.

[0021] The embodiments described in this invention are merely some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention. The term "platform" in the embodiments refers to a network content platform capable of receiving text, images, or videos and returning their publication status; "conversion" refers to a user, after browsing published content, completing pre-set business objectives such as following an account, submitting contact information, initiating an inquiry, purchasing goods, or other preset target operations; "multimodal content" includes at least text, and may also include images, audio, and video.

[0022] Example 1: As Figure 1-4As shown, this invention discloses a method for controlled publishing and feedback optimization of multimedia content, executed by a computer system. In this embodiment, the computer system is deployed on an enterprise intranet server, a cloud server, or a cluster consisting of multiple computing nodes. The operating environment in this embodiment includes: an eight-core processor, 32GB of memory, a database for storing product information, rules, and task records, object storage for storing images and videos, an inference service for running or calling a multimodal content generation model, and automated execution nodes with a browser kernel installed. The model can be called through an internal enterprise interface or run using a private deployment method.

[0023] To ensure stable data interface connectivity between different stages, the system must maintain at least the following records: Product Record (P), Content Record (C), Release Task (T), and Feedback Record (F). Example fields for each record are shown in Table 1.

[0024] Table 1 Examples of Core Data Objects Wherein, product_id is the product identifier; source_ref is the fact source identifier; content_id is the content identifier; version is the content version; platform_id is the target platform identifier; constraint_version is the platform constraint version; task_id is the task publishing identifier; and account_id is the target account identifier. In this embodiment, the fact source identifier source_ref is used to record and locate the fact source. The fact source refers to a local document, online knowledge base, or business system interface that can provide original evidence for product name, product model, specifications, functional information, applicable scenarios, applicable scope, and brand information, as well as the specific location of the corresponding fact content within the local document, online knowledge base, or business system interface. When the input is a PDF or Word document, source_ref consists of the file summary value, page number, and paragraph number; when the input is an Excel file, it consists of the worksheet name and row / column coordinates; when the input is an online knowledge base, it can consist of the document identifier, block identifier, and version number; and when the input is a business interface, it can consist of the interface name, response field path, and response time. This identifier allows the auditing module to access the original data upon which the statements were based, rather than relying on model memory.

[0025] In this embodiment, the system includes a data parsing module, a content generation module, a dual verification module, a task control module, a platform execution module, and a feedback optimization module. Each module can run as a functional unit within the same program, or it can be encapsulated as an independent service that communicates via a message queue or application programming interface.

[0026] The data parsing module includes a format recognizer, text and table extractor, field mapper, fact indexer, and product analysis agent. The format recognizer selects the parsing method based on file extensions, media type MIME, or interface description. The field mapper maps fields from different sources but with the same semantics to a unified field, for example, mapping model and product model to product_model. The fact indexer generates source_ref for field values ​​and saves a summary of the original text fragment for subsequent verification. The product analysis agent first assists the field mapper in identifying product fields and forming structured product data. The content generation module includes a content generation agent and a multimodal content generation model. The content generation agent organizes model input and calls the multimodal content generation model to generate platform content data. The product analysis agent extracts core functions, applicable scenarios, target audiences, and specifications that cannot be expanded from the structured product data, forming product analysis results provided to the content generation agent. The content generation agent reads the currently effective generation control version (control_version) based on the platform_id. It then organizes the product analysis results, platform constraint data, and the corresponding content structure parameters, media combination parameters, content length parameters, and release time parameters into the model input, and calls the multimodal content generation model to generate platform content data. During the initial generation, a preset initial generation control version is read. After the feedback optimization module generates a new generation control version, the new version is read during subsequent steps (S2), and the multimodal content generation model is called to generate the title, body text, tags, cover description, accompanying images, or video storyboards. The model output is first written to the candidate content area and cannot bypass double verification to directly enter the platform execution module.

[0027] The dual verification module includes a compliance review agent and a release executability verifier: the compliance review agent performs semantic and rule review; the release executability verifier verifies whether the title and body length, media file format and size, required fields, target page elements, and account capabilities meet the automatic execution conditions. The task control module includes a task scheduling agent, a content state machine, and a task state machine. It maintains content and task states, generates task_ids, and determines whether to call the content generation module, dual verification module, or platform execution module based on each state. The task scheduling agent determines whether to call the content generation module, dual verification module, or platform execution module based on the READY, REVISION, or BLOCKED state output by the content state machine and the execution state recorded by the task state machine. The platform execution module loads specific platform plugins through a unified plugin interface. Each plugin must implement at least the interfaces for session checking, page pre-inspection, content filling, media uploading, submission, result confirmation, and feedback reading. The feedback optimization module performs caliber conversion, normalization, and scoring on feedback metrics to generate generation control data for the next round.

[0028] The embodiments of the present invention are performed according to the following steps.

[0029] S1, Obtain Product Source Data. Users select local document data on the operations task interface, including but not limited to PDF, Word, Excel, TXT, and Markdown files, or configure Lark Docs, Yuque, enterprise internal knowledge base, and business system interface data. The data parsing module calculates the input version summary and establishes a collection batch number; if the summary from the same source remains unchanged, the verified parsing results can be reused; if changes occur, a new product data version is created to avoid referencing outdated parameters in older content.

[0030] Structured product data is generated. The data parsing module first extracts paragraphs, table cells, image descriptions, and interface fields. Then, it uses a field dictionary and a product analysis agent to identify product names, specifications, functional information, applicable scenarios, brand information, target audience, and prohibited expressions, and associates each field with a fact source identifier. For conflicting sources, the system selects the primary value based on preset source priority and writes the remaining values ​​into a conflict list for confirmation; unconfirmed conflicting fields are marked as unusable for generation. Each confirmed field, along with source_ref, is written into product record P.

[0031] S2, the platform constrains data and generates candidate content. The system uses a combination of the target platform identifier (platform_id) and the platform constraint version (constraint_version) as the query key to read the platform constraint configuration corresponding to the target platform and its current rule version. This configuration includes at least the title length range, body text number or duration range, paragraph or topic structure, image aspect ratio and size, video encoding and container format, required page fields, available publishing time window, and publishing completion criteria. When the target platform's text length, media specifications, page elements, or publishing completion criteria change, the system generates a new constraint_version while retaining the original platform constraint version for tracing already generated or published content. Page element constraints describe the page elements that the platform execution module needs to identify and manipulate on the target publishing page, including title input boxes, body text editors, topic tag input areas, media upload controls, submission controls, and publishing status prompt areas. The positioning information for each page element includes at least one of the following: element tag, element attributes, Cascading Style Sheets (CSS) selectors, XPath path positioning expressions, and adjacent tag relationships. The operable states of page elements include at least one of the following: visible, editable, uploadable, and submitable states.

[0032] The content generation agent organizes at least one of the following according to the content structure constraints: title, body text, topic tags, images, and videos, and converts this configuration into generation instructions executable by the model. For example, for text and image platforms, the output object may include title, body, topics, cover_image, and gallery; for short video platforms, the output object may include script, shot_list, voice_text, video_file, and cover_image. Here, title represents the title field, body represents the body text field, topics represents the topic tags field, cover_image represents the cover image field, gallery represents the image collection field; script represents the video script field, shot_list represents the shot list field, voice_text represents the narration text field, and video_file represents the video file field.

[0033] During the generation process, product fact fields are provided to the model in the form of "field value - source_ref" pairs, while content structure parameters and brand information are adjustable variables, and prohibited items are hard constraints. After model generation, the content generation module also performs deterministic post-processing: unifying character encoding, removing invisible characters, converting media to a resolution and container format allowed by the platform, and calculating file digest values. This yields candidate platform content data carrying content_id, version, platform_id, and constraint_version. Content_id is the content identifier, used to uniquely identify platform content data generated for a target platform and its subsequent revised versions; version is the content version, used to distinguish different generated results under the same content_id. When candidate platform content data is generated for the first time, version is set to the initial version; after each revision, content_id remains unchanged while version is incremented. The previous content version is converted to a read-only history version; only platform content data of the current version that passes both verifications can generate a release task.

[0034] S3, Perform content compliance verification. In this embodiment, content constraint data is stored in a versioned manner, including target platform content rules, legal compliance rules, brand guidelines, and prohibited items. Each rule records at least the rule identifier, rule content, severity, and whether automatic correction is allowed. During content compliance verification, the above rules are read according to the content constraint version corresponding to the current task. In this embodiment, factual statements refer to objective statements in candidate platform content regarding product names, models, specifications, performance values, functions, applicable scenarios, scope of application, usage conditions, or comparison results, and can be verified based on product source data. The compliance auditing agent first uses a rule engine to identify sensitive words, absolute claims, missing disclaimers, and prohibited brand expressions. Then, it performs source consistency checks on factual statements, including segmenting candidate content into sentences, extracting statements containing numerical values, comparatives, functional promises, or scope of application, and comparing them with field values ​​and the corresponding original text in the structured product data. Statements whose sources cannot be found, contradict the source, or exceed the source's limitations are marked as factual conflicts. The compliance verification report (compliance_report) is composed of the results from the rules engine, the fact verification results, and the model semantic audit results.

[0035] The compliance audit agent outputs three statuses based on the compliance_report: passed, correctable failure, and uncorrectable failure. Pass indicates no validation items requiring blocking or correction were found; correctable failure indicates validation items that can be eliminated through deletion, replacement, shortening, field addition, or media conversion; uncorrectable failure indicates validation items with conflicting factual sources, violating unpublishable rules, or that cannot be eliminated through allowed modification methods. For correctable failures, the compliance audit agent writes the validation item identifier, corresponding text location, rule basis, modification instruction, and allowed modification scope into the revision_context. The threshold for the number of corrections is preset by the content constraint configuration, for example, set to 3 times; when the number of corrections reaches the threshold and there are still failed validation items, the corresponding platform content data is determined as terminated content and enters the BLOCKED state.

[0036] Perform a publish executability check. The executability validator is used to determine whether the current version can be published by the corresponding platform plugin. The validator verifies the title, body text, number of tags, image size, file size, video duration, encoding format, and required fields item by item; then it loads the target page in read-only mode to confirm that the account session is valid, the publishing entry is accessible, the required page elements exist, and the account has the corresponding publishing capability, without performing a final commit.

[0037] The system transitions between states based on the results of dual verification. If both verifications pass, the platform content data is determined to be publishable and enters the READY state. If issues exist that can be resolved through text rewriting, field completion, or media conversion, it is determined to be content requiring correction and enters the REVISION state. If factual conflicts, non-publishable rules are triggered, the account lacks permissions, or the number of corrections reaches a threshold, it is determined to be terminated and enters the BLOCKED state. For the REVISION state, the system constructs a revision_context from the issue code, issue location, rule basis, allowed modification range, and target format. The content to be corrected, the revision_context, and the current content version are then input into the multimodal content generation model. The content generation agent only modifies the corresponding segments and increments the version number. The corrected platform content data undergoes both verifications again and does not inherit the passing status of the previous version. Factual fields and prohibited items not specified by the revision_context remain unchanged.

[0038] In one embodiment, a platform limits the title to no more than 20 Chinese characters, while the candidate title has 24 Chinese characters. The executability validator outputs "TITLE_LENGTH_EXCEEDED" (title length exceeds the limit), the current length, the target length, and the keywords to be retained. The system shortens the title accordingly. If the body text contains the expression "100% removal" which is not recorded in the product information, the compliance review agent outputs "FACT_UNSUPPORTED" (factual statement lacks source support) and the corresponding source_ref missing information. The system deletes the expression instead of supplementing the proof out of thin air. If an absolute term prohibited by law is triggered and the rule is configured to be uncorrectable and blocking, the system directly enters the BLOCKED state.

[0039] S4, Generate Publishing Task. The task control module binds the content_id, version, platform_id, and account_id of the publishable content to form a publishing task, and generates a task_id based on the above fields. For example, the above fields are concatenated in a preset order and then the digest is calculated using the SHA-256 secure hash algorithm, and the result is used as the task_id. Only the current content version corresponding to the READY state can generate a publishing task.

[0040] S5 executes the publishing task and determines the publishing result. The platform execution module calls the platform plugin corresponding to platform_id, first reading the publishing task snapshot and verifying the media file digest value; then, it restores the login session corresponding to account_id in the browser user directory isolated from other accounts, opens the publishing page, fills in the title, body, and tags according to the platform element mapping, uploads the media file, waits for the platform to complete the upload signal, and then executes the submission. By waiting for page state changes instead of sleeping for a fixed period of time, it can adapt to differences in network and media size.

[0041] After the plugin is submitted, it checks whether the target page displays a preset publication completion flag, determines whether the current page address matches the address pattern of the publication result page, and queries the platform's open interface or publication record page to see if a remote publication record identifier corresponding to the task_id has been generated. Once the page displays the publication completion flag, a publication result address is generated, and a remote publication record identifier is found, publication is confirmed as successful, and the publication status information SUCCESS is generated. The platform execution module then captures the publication result page, associating and saving the publication status evidence, result screenshot, task identifier, content identifier, and content version as audit evidence for the publication result. The platform execution module detects anomalies at each execution stage—session checking, page pre-inspection, content filling, media uploading, submission, and result confirmation—and outputs anomaly information including the anomaly code, the stage of occurrence, page status, retry suggestion, and on-site record. The task control module receives the anomaly information, determines whether the execution anomaly is recoverable or unrecoverable based on the anomaly code and the stage of occurrence, and executes the handling process of retrying after querying the publication record or terminating the task accordingly. Network timeouts, temporary connection interruptions, and page loading failures are configured as recoverable exceptions; invalid login status, CAPTCHAs requiring manual processing, content explicitly rejected by the platform, and unrecognizable key page elements are determined as unrecoverable exceptions.

[0042] For recoverable exceptions, the task control module first queries the local execution record using the task_id, content_id, and version, and then calls the plugin's release record query interface. If a successful record for the same version already exists, it does not resubmit but completes the local status; if no successful record exists, it retryes at backoff intervals of 30 seconds, 2 minutes, and 10 minutes. After the number of retries reaches the threshold, it is transferred to manual handling. For unrecoverable exceptions, the system terminates the task and records the exception type, page address, element location information, screenshot address, and time, for maintenance personnel to update sessions or platform plugins.

[0043] S6, the feedback optimization module performs caliber conversion, normalization, and scoring on feedback metrics to generate generation control data for the next round. Calibration conversion here refers to mapping feedback metrics recorded using different names and statistical methods on different target platforms to unified exposure metrics, click metrics, and conversion metrics, and unifying the collection time window, unit of measurement, and cumulative statistical method for feedback data. Specifically, the number of times content is displayed or reached is mapped to exposure metrics, the number of times content is clicked or detail page is visited is mapped to click metrics, and the number of times actions such as following, submitting contact information, initiating inquiries, or making purchases are mapped to conversion metrics. For cumulative metrics, the incremental value within the current collection window is obtained based on the difference in cumulative values ​​between adjacent collection times. The feedback optimization module reads the exposure metrics E, click metrics K, and conversion metrics V of the published content according to the preset collection window and associates them with the content record using content_id, version, and platform_id. To avoid directly adding the statistical calibers of different platforms, quantile normalization is first performed within each platform to obtain e, k, v ∈ [0,1], and then the content effectiveness score Q is calculated. Quantile normalization refers to ranking the exposure, click, and conversion metrics of each content item within a comparison group consisting of the same target platform, similar product categories, and the same data collection window, and converting each metric to a value between 0 and 1 based on its ranking position within the comparison group. For a comparison group containing n items, when n is greater than 1, the normalized metric for the i-th item is its ranking number minus 1, divided by n minus 1; when there are identical metric values, the average ranking number is used; when the comparison group contains only one item, the corresponding normalized metric is set to 0.5. In one embodiment, the expression is: ,in , , and + + =1, when it is transformed into the primary objective, let ,and . , , The weights for corresponding metrics are pre-set by the task configuration data, and different weight combinations can be selected based on operational goals. When conversion is the primary goal, the conversion metric weight is no less than the exposure and click metric weights. The initial run uses preset weights, which can be adjusted in subsequent runs based on the performance evaluation results of published content. The exposure, click, and conversion metric weights are determined by the preset optimization goals of the operational task and stored in the feedback evaluation configuration. The feedback evaluation configuration is selected by users with configuration permissions when creating the task, or read by the system based on a pre-established task type template. When the task's primary goal is content reach, the exposure metric weight is increased; when the task's primary goal is user access, the click metric weight is increased; when the task's primary goal is target operation completion, the conversion metric weight is increased. When conversion is the primary goal, the weights can be adjusted accordingly. , and The values ​​were set to 0.2, 0.3, and 0.5 respectively. The weights used and their configuration versions were recorded for each scoring session to allow for tracking of the scoring results.

[0044] The feedback optimization module treats content from the same platform, similar product categories, and within the same collection window as a comparison group, analyzing the relationship between title structure, body length, media mix, and publication time and Q. For discrete parameters such as content structure and media mix, the feedback optimization module calculates the average content effectiveness score corresponding to each parameter value. When the number of valid samples for a parameter value reaches a preset sample threshold, and its average content effectiveness score is higher than the average content effectiveness score of its comparison group, the selection weight of that parameter value is increased according to a preset update step size; if it is lower than the average content effectiveness score of its comparison group, its selection weight is decreased according to a preset update step size. After adjustment, the selection weights of each value under the same parameter are normalized so that their sum is 1. For interval parameters such as content length and publication time, the parameter interval with the higher average content effectiveness score is selected as the priority value interval for the next round. For example, the preset sample threshold can be set to 10, and the update step size of the selection weight can be set to 0.05. These values ​​can be configured according to the amount of content on the target platform and the feedback cycle. In one embodiment, if the "Scenario Problem - Solution Process - Parameter Basis" structure achieves a higher Q across multiple versions, the selection weight of this structure template is increased; if a higher post-click conversion rate is consistently achieved during a certain release period, the release time parameter is adjusted. The update result is saved as a generated control version (control_version), recording the sample range, indicator weight, and effective time. control_version represents the generated control data version, used to distinguish the content structure parameters, media combination parameters, content length parameters, and release time parameters generated by the feedback optimization module in different feedback cycles. Regardless of how Q changes, the feedback optimization module must not modify the product fact value, the original text corresponding to the fact source identifier (source_ref), or prohibited items; the feedback optimization module can only modify the content structure parameters, media combination parameters, content length parameters, and release time parameters listed in the writable field whitelist. The prohibition on modification only restricts the feedback optimization module; when the product source data is updated, the data parsing module can still generate a new product data version and corresponding fact source identifier based on the updated product source data. When a parameter update request contains a field not in the writable field whitelist, the feedback optimization module refuses to execute the update of that field and records the reason for refusal. After the feedback optimization module completes the parameter update, it generates a new control_version and sets it as the current effective version for the next feedback cycle of the corresponding target platform. When the task control module starts the subsequent step S2, it provides the current effective version to the content generation module.

[0045] In one embodiment, after receiving the task "Generate and publish content for product P001 on four platforms," ​​the task scheduling agent first establishes a run instance identifier (run_id). The product analysis agent reads the structured fields of P001 and outputs the product analysis result (analysis_result), which includes core selling point candidates, applicable scenarios, target audience, parameters that must be retained, and fields that cannot be expanded. After the scheduler verifies the pattern of the analysis_result, it sends the corresponding platform constraints to the four content generation agent instances respectively, instead of simply truncating the same article before publication.

[0046] After the content generation agent outputs candidate platform content data (candidate_content), the scheduler simultaneously triggers the compliance review agent and the release executability validator. Both return reports with content_id and version to prevent incorrect binding of previous version reports to the new version. If either report fails to pass due to being remediable, the scheduler merges the issues but maintains rule priority: factual conflicts and legal rules take precedence over writing style, length, and media layout. The revision context (revision_context) received by the content generation agent only contains fields that are allowed to be modified; after modification, a new version is generated.

[0047] Once the content on the four platforms reaches the READY state, the task scheduling agent can execute independent publishing tasks in parallel. However, tasks from the same account are executed sequentially according to an account-level mutex lock to avoid multiple browser pages contending for login sessions. The status of each task is recorded as CREATED, PRECHECKED, FILLING, UPLOADING, SUBMITTING, VERIFYING, and SUCCESS or FAILED, with each status change carrying a timestamp. Therefore, even if the system restarts, unfinished tasks can be recovered based on the last persistent status. Specifically, CREATED indicates the task has been created, PRECHECKED indicates pre-check completion, FILLING indicates content filling, UPLOADING indicates media uploading, SUBMITTING indicates submission in progress, VERIFYING indicates result confirmation, SUCCESS indicates successful publishing, and FAILED indicates publishing failure.

[0048] This embodiment uses a "desktop air quality monitor" to illustrate how data flows between each step. Product source data includes the product manual (PDF), specifications (Excel), and a brand thesaurus. The data parsing module obtains the product model AQ-M1, sensor parameters, sampling period, power supply method, applicable temperature and humidity range, and prohibited brand items ("cannot claim medical diagnostics"), and establishes source_refs for each.

[0049] The system reads the constraints for text / image platform A: "title no more than 20 Chinese characters, body text 300 to 1000 characters, cover aspect ratio 3:4, and no more than 10 hashtags." For video platform B, it reads the constraints: "video length 15 to 60 seconds, cover aspect ratio 9:16, and title no more than 30 characters." The content generation agent then generates text / image drafts and video scripts accordingly. The specification values ​​in the text / image drafts all carry `source_ref`; each shot in the video script includes a scene description, narration, duration, and corresponding product facts.

[0050] During the initial validation, the title of the text and image contained 22 Chinese characters, and the executability validation result was "correctable but failed." The text contained the phrase "help diagnose air quality issues," which conflicted with the brand's prohibited items. Since this conflicting statement could be eliminated by deletion without altering other product facts, the content compliance validation result was "correctable but failed." The system generated a revision_context, requiring the title to be compressed to under 20 characters and the removal of statements with medical diagnostic implications, while keeping factual fields such as the sampling period unchanged. The revised version (version=2) passed both validations and a release task was generated.

[0051] The platform execution module restores the session of account A on platform, opens the editing page, and locates the title box, body editor, cover upload area, and topic area based on element tags and the backup CSS locator. After filling in the text, the module reads the character count displayed on the page, uploads the cover, verifies the returned media summary value, and finally submits. The page displays a publication completion mark and generates a result address. The module captures the result page and writes the three status pieces of evidence into the task record. Subsequently, it collects exposure, click, and conversion data in 24-hour and 72-hour windows, and associates them with content_id and version=2.

[0052] Once a set number of content versions for similar products have accumulated, the feedback optimization module discovers that content using a "usage scenario introduction – verifiable parameter description – operation suggestions" structure on platform A scores higher than content using a pure parameter list. Therefore, the weight of this structure template is increased. Simultaneously, it's found that versions showing changes in monitored values ​​in the first 3 seconds of the 9:16 video have higher conversion rates after clicks, so the camera structure parameters on platform B are updated. Subsequent generation can only reference confirmed facts from product records; therefore, the feedback loop will not solidify exaggerated statements with high conversion rates but lacking factual basis into generation strategies.

[0053] This embodiment conducts a simulation experiment to verify the impact of the present invention on automation stability and content effect matching. Four types of platform constraints and 50 sets of anonymized product records are set, resulting in 200 publishing tasks. These tasks include 30 pre-set content rule conflicts, 20 media specification errors, 15 network or page timeout exceptions, and 10 duplicate scheduling requests. The platform adapter returns the page state according to fixed rules and does not access the real platform. Comparison scheme A uses a linear process of "automatic publishing directly after generation"; comparison scheme B uses a process of "automatic publishing after content compliance verification"; the present invention's scheme C employs a closed loop of fact source association, dual verification, versioned tasks, idempotent confirmation, and feedback optimization.

[0054] Each scheme uses the same input data, model version, and platform constraints, runs three times consecutively, and the average value is taken. The violation or non-executable content blocking rate is the ratio of the number of abnormal content correctly blocked or corrected by the system before submission to the preset number of abnormal content; the release confirmation accuracy rate is the proportion of tasks whose recorded status matches the actual status of the simulation platform; the number of manual interventions is the average number of times each task requires further manual processing; the simulation platform outputs exposure metrics E, click metrics K, and conversion metrics V according to the same simulation feedback generation rules for each scheme, based on the matching results of content with platform templates, release time, and audience tags; the feedback optimization module jointly sorts and normalizes the quantiles of the feedback metrics obtained by the three schemes within the same target platform, similar product categories, and the same collection window, and then calculates the normalized average exposure metric, normalized average click metric, and normalized average conversion metric for all effective content in each scheme, and calculates the average content effectiveness score Q for each scheme accordingly. In this embodiment, the content effectiveness score Q is a simulation evaluation result used to compare the matching effect of generated content with target platform conditions under different schemes. Each comparison scheme uses the same simulation feedback generation rules, metric weights, and comparison group division method. Example results are shown in Table 2.

[0055] Table 2 Comparison Results of Closed Simulation As shown in Table 2, content compliance verification can improve the interception rate of abnormal content, but it still has shortcomings regarding media specifications, page elements, and repetitive tasks. By adding publish executability verification and task idempotency control, Solution C further reduces manual intervention and repetitive publishing under pre-defined conditions. The second-round simulation content effect score Q of Solution C is approximately 11.9% higher than that of Solution B. Under the simulation conditions set in this embodiment, associating the effect feedback with the corresponding content version and using it to update the generation control data can improve the matching degree between subsequent generated content and the target platform conditions.

[0056] In other embodiments, the platform execution module may not employ browser automation but instead call open interfaces provided by the target platform; as long as this module still performs publishing based on the content version and returns a verifiable publishing status, it constitutes an equivalent implementation of the platform execution module. The compliance review agent can call the rule engine to perform deterministic rule verification.

[0057] This invention can also be implemented by an electronic device. The electronic device includes one or more processors and a memory, in which a computer program is stored, and the processor executes the program to perform the above-described method. The program can also be stored in a non-transitory computer-readable storage medium. The system modules are divided logically; a module can be implemented by multiple programs or devices, and multiple modules can be integrated into the same program or device.

[0058] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for controlled publishing and feedback optimization of multimedia content, characterized in that, Includes the following steps: S1. Parse the product source data to obtain structured product data; S2. For at least one target platform, based on the structured product data and the platform constraint data corresponding to the target platform, generate platform content data through a multimodal content generation model; S3. Perform content compliance verification on the platform content data based on the content constraint data, and perform publication executability verification on the platform content data based on the platform constraint data. Based on the results of the two verifications, determine the platform content data as publishable content, content to be corrected, or terminated content; wherein, the content to be corrected is returned to the multimodal content generation model for correction, and the terminated content is blocked. S4. Bind the content identifier and content version in the publishable content to the target platform to generate a publishing task; S5. The platform execution module executes the publishing task and determines the publishing result based on the publishing status information returned by the target platform; S6. Collect the effect feedback data corresponding to the successfully published content, associate the effect feedback data with the publishable content according to the content identifier to obtain the effect evaluation result, and update the generation control data of subsequent platform content data according to the effect evaluation result under the condition of satisfying the content constraint data and the platform constraint data.

2. The multimedia content controlled publishing and feedback optimization method according to claim 1, characterized in that, In step S1, the product source data includes at least one of local document data, online knowledge base data, and business system interface data; the structured product data includes at least two of product name, specifications, functional information, applicable scenarios, and brand information, and each field is associated with a fact source identifier, which is used to verify the consistency between factual statements and product source data during content compliance verification.

3. The multimedia content controlled publishing and feedback optimization method according to claim 1, characterized in that, The platform constraint data mentioned in step S2 includes text length constraints, content structure constraints, media specification constraints, page element constraints, and publication completion judgment conditions; when generating the platform content data, at least one of the following is organized according to the content structure constraints: title, body text, topic tags, images, and videos, and corresponding media files are generated and converted according to the media specification constraints.

4. The method for controlled publishing and feedback optimization of multimedia content according to claim 1, characterized in that, The method is implemented through a product analysis intelligent agent, a content generation intelligent agent, a compliance review intelligent agent, and a task scheduling intelligent agent. The product analysis intelligent agent is used to assist in generating the structured product data, the content generation intelligent agent is used to call a multimodal content generation model to generate the platform content data, the compliance review intelligent agent is used to perform content compliance verification, and the task scheduling intelligent agent calls the platform execution module to execute the publishing task.

5. The multimedia content controlled publishing and feedback optimization method according to claim 1, characterized in that, In step S3, the content constraint data includes target platform rules, legal compliance rules, and brand specifications; the content compliance verification outputs three states: pass, correctable failure, and uncorrectable failure. The correctable failure verification items, their corresponding text positions, and modification instructions are written into the correction context, and the correction context and the content to be corrected are input into the multimodal content generation model. If the verification still fails after the number of corrections reaches a threshold, the corresponding content is determined as terminated content.

6. The multimedia content controlled publishing and feedback optimization method according to claim 1, characterized in that, Assign the content identifier and the corrected content version to each platform's content data, and generate a task identifier based on the content identifier, content version, target platform, and target account; Before resubmitting, the publishing record is queried based on the task identifier. The resubmission operation is only performed if there is no successful publishing record.

7. The method for controlled publishing and feedback optimization of multimedia content according to claim 1, characterized in that, The platform execution module connects to different target platforms. It automatically loads isolated account sessions that are still logged in, locates page elements on the target platform, performs text filling, media uploading, and submission operations, and generates publishing status information indicating successful publishing when at least two of the following indicate successful publishing: publishing completion flag, publishing result address, and remote publishing record identifier. It also generates a screenshot of the publishing result and saves it in association with the publishing status information.

8. The method for controlled publishing and feedback optimization of multimedia content according to claim 1, characterized in that, In step S6, the effect feedback data includes exposure, clicks, and conversions. Based on the statistics of each target platform, the effect feedback data is normalized to obtain a content effect score. At least one of the content structure parameters, media combination parameters, content length parameters, and release time parameters in the generation control data is adjusted according to the content effect score. At the same time, the fact fields in the structured product data and the prohibited items in the content constraint data are set as non-writable fields of the feedback optimization module, so that the feedback optimization module only updates the adjustable parameters in the generation control data.

9. A controlled publishing and feedback optimization system for multimedia content, characterized in that, include: The data parsing module is equipped with a product analysis intelligent agent, which is used to parse the product source data to obtain structured product data; The content generation module is equipped with a content generation intelligent agent, which generates platform content data based on the structured product data and the platform constraint data corresponding to the target platform through a multimodal content generation model. The dual verification module is equipped with a compliance review intelligent agent and a release executability verifier. The compliance review intelligent agent is used to perform content compliance verification, and the release executability verifier is used to perform release executability verification. Based on the results of the two verifications, the module determines whether the content is publishable, content to be corrected, or content to be terminated. The task control module is equipped with a task scheduling agent to bind the content identifier and content version of the publishable content to the target platform to generate a publishing task; The platform execution module is used to execute the publishing task and determine the publishing result based on the publishing status information of the target platform; The feedback optimization module is used to associate the effect feedback data of successfully published content with the content identifier to obtain the effect evaluation result, and update the subsequent generation control data under the condition of satisfying the content constraint data and the platform constraint data.

10. The system according to claim 9, characterized in that, The platform execution module is used to detect execution anomalies during the execution of the publishing task and output anomaly information to the task control module; The task control module is used to determine whether the execution exception is a recoverable exception or an unrecoverable exception based on the exception information. For recoverable anomalies, after confirming that there is no successful publication record based on the task identifier, the platform execution module is triggered to re-execute the corresponding publication task according to the preset backoff interval; for unrecoverable anomalies, the corresponding publication task is terminated, and the task identifier, target platform, anomaly type, page status, screenshot address, and occurrence time are recorded.

Images