A media platform-oriented risk control-driven RPA automation publishing method and system

Abstract

The application discloses a media platform-oriented risk control driven RPA automation publishing method and system, and belongs to the technical field of RPA and AIGC. The system comprises a content generation and semantic security layer, a risk control and compliance engine, an arrangement and dynamic regulation layer, a full-link evidence chain module, a regional combination module, a monitoring and early warning module and a visual management and control console. The core is driven by risk control, R-score and C-score are calculated through multi-model weighted fusion, and RPA task strategies are dynamically regulated. The method comprises the steps of content generation, risk control evaluation, task scheduling, evidence chain storage, monitoring and early warning, and rule updating. The application solves the defects of process interruption, disconnection between risk control and execution, and lack of full-link traceability in the prior art, improves the stability, compliance and traceability of new media automation publishing, reduces maintenance cost and account risk, adapts to multi-region and multi-platform requirements, and can realize commercialization.

Description

Technical Field

[0001] This invention belongs to the field of robotic process automation technology, specifically relating to a risk control-driven RPA automated publishing method and system for media platforms. Background Technology

[0002] With the rapid development of the new media industry, the demand for large-scale and automated content production and distribution is becoming increasingly urgent. The combination of RPA (Robotic Process Automation) and AIGC (Artificial Intelligence Generated Content) has become the mainstream solution in the industry, enabling functions such as automatic content generation, batch publishing across multiple platforms, and scheduled task execution, significantly improving content production efficiency. However, in practical implementation, existing RPA+AIGC new media automated publishing technologies suffer from the following insurmountable technical defects due to the characteristics of media platforms, compliance requirements, and technical architecture limitations. These defects result in poor solution stability, high maintenance costs, and significant account risks, making it difficult to meet the needs of large-scale commercial applications: New media platforms frequently update their UI. Existing RPA technologies are mostly based on fixed DOM paths for control location. Once the platform's UI layout, button positions, or page structure change, the location logic immediately fails, causing the RPA process to be frequently interrupted and unable to execute normally. This requires continuous manual maintenance of the location rules, which is extremely costly. During operation, various abnormal scenarios are prone to occur, such as platform pop-up prompts, page loading timeouts, account login failures, content review pop-ups, etc. Existing RPA has no adaptive exception handling capabilities. In the face of such scenarios, it will directly freeze, miss content, or publish repeatedly, further aggravating the platform's violation risks. There is no mechanism to resume the process after it is interrupted. The entire publishing process must be re-executed from the beginning, which can easily lead to the same content being published repeatedly or the content publishing order being disordered. This not only affects the user experience, but may also trigger platform traffic restrictions and penalties. Content risk control and RPA execution are completely disconnected. AIGC-generated content has not undergone strict compliance testing, or the testing standards do not match the compliance requirements of the target platform and target region. High-risk content is directly and automatically published through RPA, which can easily trigger serious consequences such as platform bans and account association penalties. When multiple accounts publish in batches, the existing solution lacks a reasonable resource scheduling and environment isolation mechanism, resulting in problems such as resource contention between accounts, mechanical operation behavior (such as completely consistent publishing time and operation frequency), and lack of isolation of the operating environment. The platform can easily identify it as machine operation, and then ban multiple accounts together, causing significant economic losses.

[0003] To address the aforementioned technical shortcomings, there is currently no effective solution in the industry. There is an urgent need for an automated publishing system and method that takes risk control as its core, links RPA and AIGC, and has the capabilities of adaptive anomaly handling, breakpoint resumption, full-link traceability, and regional compliance adaptation. This would solve the pain points of existing solutions and promote the large-scale commercial application of RPA+AIGC technology in the new media field. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a risk control-driven RPA automated publishing method and system for media platforms, thereby solving the aforementioned technical problems in the prior art.

[0005] The objective of this invention can be achieved through the following technical solutions: A risk control-driven RPA automated publishing system for media platforms, comprising: The content generation and semantic security layer integrates a Prompt closed-loop optimization algorithm, which is used to automatically rewrite the Prompt based on the type of risk control violation and iteratively generate compliant content through reinforcement learning. The risk control and compliance engine calculates R-score and C-score through a multi-model weighted fusion algorithm, and outputs risk level and regional compliance recommendations. The orchestration and dynamic control layer, based on the risk control-RPA linkage algorithm, adjusts the priority, concurrency, release channels and resource allocation strategies of RPA tasks in real time according to the R-score. When the R-score > 60, the task is automatically paused and a full-link rollback is triggered. The full-link evidence chain module, based on the SHA-256 hash chain and distributed ledger storage algorithm, links the Prompt version ID, RPA task ID, risk control results, release records and user operation logs into an immutable full-link evidence chain. The regional compliance module automatically crawls and extracts compliance rules for the target region / platform through rule parsing and adaptive algorithms, and dynamically adjusts content generation and review strategies. The monitoring and early warning module collects full-link metrics in real time. When the violation rate or rollback rate exceeds the threshold, it triggers automatic early warning and strategy optimization.

[0006] Furthermore, the Prompt closed-loop optimization algorithm includes a reinforcement learning iterative mechanism, specifically: Construct the reward function: Reward = (1 - probability of violation) × 0.7 + content quality score × 0.3; The content quality score is calculated by weighting three dimensions: readability, relevance, and originality. When the compliance rate of the rewritten content is >95%, the weight of the corresponding template increases by 10%. When the compliance rate is less than 60%, the template weight is reduced by 20% and manual review is triggered; The template library is automatically updated every 500 iterations, adding high-quality templates and removing low-quality ones.

[0007] Furthermore, the multi-model weighted fusion algorithm of the risk control and compliance engine is specifically as follows: Multi-model score = BERT sensitive word matching score × 0.3 + GPT multimodal detection score × 0.4 + localization rule matching score × 0.3; Among them, the regional rule matching score is calculated based on the rule base built on the laws and regulations of the target region and the platform specifications, and is updated in real time.

[0008] Furthermore, the risk control-RPA linkage algorithm of the orchestration and dynamic control layer also includes a dynamic resource scheduling strategy: When R-score < 30, allocate 80% of the RPA resource pool for high-concurrency deployments; When 30≤R-score<60, allocate 40% of the resource pool and restrict tasks to be executed only on channels with a compliance rating of ≥B; When R-score > 60, release all resources occupied by the task and trigger resource reflow, prioritizing allocation to low-risk tasks.

[0009] Furthermore, the rule parsing and adaptive algorithm of the regionalized compliance module includes: It automatically crawls the target platform's rule update documents every week, parses the rule text using the LLM algorithm, and extracts sensitive content keywords, compliance restrictions, and penalty rules. Construct a regional rule feature library, with feature dimensions including: political sensitivity index, copyright strictness, and tolerance for vulgar content; Based on the target area of ​​the RPA task, dynamically adjust the Prompt constraints and risk control detection thresholds for content generation.

[0010] Furthermore, the SHA-256 hash chain + distributed ledger storage algorithm of the full-link evidence chain module is specifically as follows: A three-level hash chain is generated: content generation hash, risk control review hash, and release execution hash. Each level of hash is calculated based on the previous level of hash, forming an immutable association. It adopts a lightweight distributed ledger storage hash chain, which supports tracing the entire chain of information by any of the identifiers in the Prompt version ID, RPA task ID, and published content hash; The evidence chain query interface supports exporting compliance audit reports, which include the entire task process logs and hash verification information.

[0011] Furthermore, the monitoring and early warning module includes: Real-time collection of 12 core metrics: violation rate, rollback rate, success rate of Prompt rewriting, resource utilization rate, changes in channel compliance rating, frequency of regional rule updates, manual intervention rate, content production efficiency, R-score distribution, C-score distribution, task timeout rate, and user complaint rate. When any metric exceeds a preset threshold, a multi-channel alert is triggered, and strategy optimization suggestions are automatically generated. It supports customizable monitoring panels, allowing users to configure key metrics and alert thresholds according to their needs.

[0012] Furthermore, it also includes a visual management console with the following functions: Real-time risk control dashboard: Displays R-score / C-score distribution, violation type statistics, and changes in channel compliance rating; RPA task orchestration view: Supports drag-and-drop adjustment of task flow, viewing task status and resource usage; Evidence chain query module: Enter the task ID or content hash to export a full-chain audit report with one click; Rule management interface: Manually edit compliance rules and upload custom Prompt rewrite templates.

[0013] A method for deploying a risk-controlled, RPA-driven automated deployment system for media platforms includes the following steps: S1. Receive the content production request input by the user, extract the topic, target area, and publishing channel context information, and generate an initial Prompt; S2. Call the AIGC model to generate content, and calculate the R-score and C-score using a multi-model weighted fusion algorithm; S3. Adjust RPA task strategies based on the risk control-RPA linkage algorithm: If R-score < 30, allocate 80% of the resource pool to start a high-concurrency deployment; If 30 ≤ R-score < 60, allocate 40% of the resource pool and limit release to compliant channels; If the R-score is greater than 60, a Prompt closed-loop optimization is triggered. If the rewrite still violates the rules after 3 consecutive rewrites, a manual review is triggered. S4. After the release is completed, a three-level SHA-256 hash chain is generated and stored in the distributed ledger; S5. Real-time monitoring of core metrics; when a metric exceeds a threshold, an alert is triggered and optimization suggestions are generated. S6. Automatically crawl platform rules updates weekly, and update the regional rule feature library through rule parsing and adaptive algorithms.

[0014] Furthermore, the Prompt closed-loop optimization in S3 includes: The risk control engine identifies the type of violation and retrieves the corresponding rewritten template from the template library. Generate an optimized Prompt, call AIGC to regenerate the content and calculate a new R-score; The template weights are updated based on the reinforcement learning reward function, and the subsequent rewriting logic is optimized.

[0015] The beneficial effects of this invention are: 1. This invention takes risk control as the core driver and realizes deep linkage between RPA and AIGC, which solves the problem of disconnect between risk control and execution in the existing technology. By dynamically adjusting the resource allocation, concurrency and publishing channels of RPA tasks through R-score, it effectively reduces the probability of publishing high-risk content, avoids platform traffic restrictions and account bans, improves account security and reduces operational risks.

[0016] 2. This invention integrates a closed-loop optimization algorithm for Prompt with a reinforcement learning iterative mechanism, which can automatically rewrite the Prompt based on the type of risk control violation, dynamically update the template library, improve the content compliance rate, and at the same time take into account the content quality, reduce the cost of manual review, and realize the automation and compliance of content generation.

[0017] 3. This invention sets up an orchestration and dynamic control layer, which has breakpoint resume and full-link rollback mechanisms. It can adaptively handle abnormal scenarios such as pop-ups, loading timeouts, and login failures, avoid re-execution after process interruption, reduce duplicate releases and missed releases, improve RPA process stability, and reduce maintenance costs.

[0018] 4. This invention achieves automatic crawling, parsing and adaptation of compliance rules for target platforms / regions through a regional compliance module, dynamically adjusts content generation and review strategies, adapts to compliance requirements of multiple regions and platforms, expands the system's applicability, and meets global publishing needs.

[0019] 5. This invention is based on the SHA-256 hash chain and distributed ledger storage algorithm to construct an immutable end-to-end evidence chain, enabling traceability of the entire process from content generation and risk control review to RPA release. It can quickly export compliance audit reports to meet the industry's compliance audit requirements. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0021] Figure 1 This is a block diagram of an automated publishing system according to an embodiment of the present invention; Figure 2This is a schematic diagram of the method flow of the automated publishing system according to an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, this invention provides a risk control-driven RPA automated publishing system for media platforms. The overall system architecture has a layered structure, specifically including: Content generation and semantic security layer, risk control and compliance engine layer, orchestration and dynamic adjustment layer, full-link evidence chain layer, regional compliance module, monitoring and early warning layer, visual control console and rule management console.

[0024] Data flow and interaction: User request > Initial Prompt generation > Content generation > Risk control assessment > Orchestration and scheduling > RPA execution > Evidence chain implementation > Monitoring and early warning > Regional rule update.

[0025] Key feature mapping claims include: providing risk control-driven dynamic resource scheduling (R-score triggers different resource allocations), closed-loop optimization and template adaptation of Prompts, adaptive acquisition and application of regional compounding rules, immutability of the entire evidence chain, breakpoint resume and rollback mechanisms, and linkage between AI-generated content and the risk control engine.

[0026] The main functions of the content generation and semantic security layer are: It receives inputs such as topic, region, and channel, and generates an initial Prompt; it calls a multimodal / text AIGC model to output a draft; it evaluates the readability, relevance, and originality of the output text to form a content quality score; and it performs closed-loop optimization of the Prompt based on the risk control prompt template library, rewriting the output to improve compliance.

[0027] Key algorithm: Reinforcement learning-driven Prompt closed-loop optimization; Example reward function: Reward = (1 - probability of violation) × 0.7 + content quality score × 0.3; When the compliance rate after rewriting is >95%, the template weight is increased by +10%; when the compliance rate is <60%, the template weight is reduced and manual review is triggered; the template library is automatically updated every 500 iterations, eliminating low-performance templates and adding high-performance templates; rollback points are set at text generation and rewriting nodes, and if an exception occurs, it can continue from the most recent rollback point.

[0028] Risk control and compliance engine: It weights and fuses the scores of multiple models to output R-score and C-score, as well as regional compliance suggestions.

[0029] Multi-model weighted fusion formula: Match score = BERT sensitive word matching score × 0.3 + GPT multimodal detection score × 0.4 + regionalization rule matching score × 0.3; Regionalized rule base: Real-time / periodic updates of regional regulations, sensitive word lists, penalty rules, etc.; LLM is used to extract keywords and constraints.

[0030] Process control: R-score thresholds are mapped to orchestration strategies.

[0031] Orchestration and dynamic control layer: Dynamically allocate resources, adjust concurrency, limit channels, and trigger rollback based on R-score; supports breakpoint resume and full-link rollback.

[0032] Dynamic resource scheduling strategy: R-score < 30: Allocate 80% of the resource pool for high-concurrency deployments; 30≤R-score<60: Allocate 40% of the resource pool, and only implement this through compliant channels; R-score ≥ 60: Release resources and prioritize allocation to low-risk tasks; trigger manual review if continuous rewriting still violates regulations.

[0033] Key nodes record rollback points, and in case of anomalies, execution continues from the most recent rollback point; rollback point metadata is written into the evidence chain.

[0034] End-to-end evidence chain module: Constructs an immutable end-to-end evidence chain, facilitating compliance auditing and traceability.

[0035] Hash chain structure: Three-level hash chain: content generation hash (H1) → risk control review hash (H2) → release execution hash (H3), with each level of hash calculated using the previous level's hash as input.

[0036] Storage and Traceability: It adopts a lightweight distributed ledger storage and supports tracing the entire chain of information by fields such as Prompt version ID, RPA task ID, and content hash.

[0037] Regionalized compliance module: Periodically captures regional rule texts, uses LLM / LLM to parse the rule texts, and extracts sensitive keywords, compliance restrictions, penalty rules, etc.; builds a regional rule feature library, with feature dimensions including political sensitivity index, copyright strictness, tolerance for vulgar content, etc., and dynamically adjusts the prompt constraints and risk control detection thresholds for content generation.

[0038] Real-time integration with FCE means that regional rules directly affect the regional rule matching score.

[0039] Monitoring and early warning module: Real-time collection of core indicators (such as violation rate, rollback rate, success rate of Prompt rewriting, resource utilization rate, changes in channel compliance rating, frequency of regional rule updates, manual intervention rate, content production efficiency, R-score distribution, C-score distribution, task timeout rate, and user complaint rate).

[0040] When any metric exceeds a threshold, multi-channel alerts are triggered, and strategy optimization suggestions are automatically generated. Custom monitoring dashboards are also supported, allowing users to configure monitored metrics and thresholds as needed.

[0041] Visualized control console and rule management console: Risk control dashboard: R-score / C-score distribution, violation type statistics, and changes in channel compliance rating; RPA task orchestration view: drag-and-drop task workflow arrangement, task status and resource usage viewing; Evidence chain query module: Enter task ID / content hash to export audit report with one click; Rule management interface: rule editing and uploading custom Prompt rewriting templates.

[0042] like Figure 2 As shown, this relates to a method for automated dynamic management and traceable publishing of new media based on RPA and AIGC, including the following steps: S1. Receive the content production request input by the user, extract the topic, target area, and publishing channel context information, and generate an initial Prompt; S2. Call the AIGC model to generate content, and calculate the R-score and C-score using a multi-model weighted fusion algorithm; S3. Adjust RPA task strategies based on the risk control-RPA linkage algorithm: If R-score < 30, allocate 80% of the resource pool to start a high-concurrency deployment; If 30 ≤ R-score < 60, allocate 40% of the resource pool and limit release to compliant channels; If the R-score is greater than 60, a Prompt closed-loop optimization is triggered. If the rewrite still violates the rules after 3 consecutive rewrites, a manual review is triggered. S4. After the release is completed, a three-level SHA-256 hash chain is generated and stored in the distributed ledger; S5. Real-time monitoring of core metrics; when a metric exceeds a threshold, an alert is triggered and optimization suggestions are generated. S6. Automatically crawl platform rules updates weekly, and update the regional rule feature library through rule parsing and adaptive algorithms.

[0043] Example 1: Suitable for MCN agencies to publish short videos in bulk across multiple platforms This embodiment addresses the bulk publishing needs of MCN agencies across multiple accounts, platforms, and regions for short videos. It incorporates the compliance requirements of mainstream media platforms such as Douyin, Kuaishou, and TikTok to automate the generation, compliance review, bulk publishing, and end-to-end traceability of short video content. The specific implementation process is as follows: 1. System Deployment and Initialization Deploy the system described in this invention, configure an RPA resource pool (total capacity of 100 concurrent task nodes), input 50 Douyin accounts, 30 Kuaishou accounts, and 20 TikTok accounts (covering three regions: China, Southeast Asia, and Europe and America) under an MCN agency, and complete the account environment isolation configuration (each account corresponds to an independent operating environment to avoid associated detection); through the regional compliance module, initialize the compliance rules of the target platform and region, automatically crawl the domestic compliance rules of Douyin and Kuaishou (such as sensitive word lists and copyright requirements), and the regional rules of TikTok in Southeast Asia (Indonesia and Malaysia) and Europe and America (such as restrictions on religiously sensitive content and advertising compliance requirements), and build an initial regional rule feature library.

[0044] 2. Content generation and semantic security optimization Users input content production requests through the visual management console: the topic is "Spring Beauty Product Recommendations", the target regions are China (Douyin, Kuaishou) and Indonesia (TikTok), and the publishing channels are the corresponding accounts on the above three platforms. An initial Prompt is generated: "Recommending 3 popular spring beauty products, with lively language, suitable for young women, highlighting the product advantages."

[0045] The content generation and semantic security layer calls the AIGC multimodal model to generate three short video scripts and accompanying visual descriptions. Subsequently, compliance checks are performed using the Prompt closed-loop optimization algorithm. It was found that the scripts for the Indonesian region contained medical beauty-related terms such as "whitening" and "freckle removal" (violating TikTok's compliance rules for Indonesia), triggering a Prompt rewrite.

[0046] The risk control engine identified the violation type as "regional sensitive word violation". It retrieved the Indonesian beauty category Prompt from the template library and rewrote the template. The optimized Prompt is "Recommended 3 popular spring beauty products, with lively language, suitable for young women, highlighting the moisturizing and brightening effects of the products, and not involving medical beauty-related descriptions". The AIGC-generated content was re-invoked, achieving a compliance rate of 98%, corresponding to a 10% increase in template weight. The content quality score was calculated by weighting readability (85 points), relevance (90 points), and originality (88 points) (each with a weight of 1 / 3), resulting in a score of 87.67. The reward function is as follows: Reward = (1-0)×0.7+87.67 / 100×0.3=0.7+0.263=0.963, the optimization effect meets the standard.

[0047] 3. Risk control assessment and task scheduling The risk control and compliance engine calculates the R-score and C-score using a multi-model weighted fusion algorithm: BERT sensitive word matching score 92 (no sensitive words), GPT multimodal detection score 88 (no illegal elements in the image), and regional rule matching score 95 (optimized text conforms to the rules of each region). The multi-model score = 92×0.3+88×0.4+95×0.3=27.6+35.2+28.5=91.3 points, corresponding to an R-score of 100-91.3=8.7 points (the lower the R-score, the lower the risk) and a C-score of 91.3 points (the higher the compliance score, the better the compliance). The output risk level is "low risk", and the regional compliance suggestion is "can be published normally, but pay attention to avoiding religious elements in the TikTok Indonesia region".

[0048] The orchestration and dynamic control layer, based on R-score (8.7 < 30), executes a dynamic resource scheduling strategy, allocates 80% of the RPA resource pool (80 concurrent nodes), and initiates high-concurrency deployment. The deployment tasks of 50 Douyin accounts, 30 Kuaishou accounts, and 20 TikTok accounts are prioritized (Douyin and Kuaishou first, followed by TikTok Indonesia region), and a concurrent deployment limit is set (20 per batch for Douyin, 15 per batch for Kuaishou, and 10 per batch for TikTok) to avoid the platform detecting abnormal concurrency.

[0049] 4. RPA Execution and Exception Handling After the RPA task is started, it automatically completes operations such as account login (using passwordless login + automatic CAPTCHA recognition to avoid login failure), short video upload, title / topic setting, and posting time calibration. During the process, one TikTok Indonesian account encountered a "content review pop-up" anomaly. The system automatically triggered adaptive anomaly handling, suspending the posting task for that account, recording the breakpoint information (current progress is "upload completed, awaiting review"), and releasing the resources occupied by that account to allocate to other low-risk tasks. After the review is approved, the system continues to execute the posting process from the breakpoint to avoid repeating the process from the beginning and causing duplicate posting.

[0050] 5. Full-chain evidence and monitoring / early warning After the release is completed, the end-to-end evidence chain module generates a three-level hash chain: content generation hash (H1, calculated based on the initial Prompt, optimized Prompt, and AIGC-generated content), risk control review hash (H2, calculated based on R-score, C-score, and risk control results), and release execution hash (H3, calculated based on the releasing account, release time, and release result). H2 is calculated based on H1, and H3 is calculated based on H2, forming an immutable association. The hash chain is stored in a lightweight distributed ledger and associated with the Prompt version ID, RPA task ID, release record, and user operation log.

[0051] The monitoring and early warning module collects 12 core indicators in real time. During this release, the violation rate was 0%, the rollback rate was 0%, the prompt rewrite success rate was 100%, and the resource utilization rate was 78% (not exceeding the threshold of 80%). There were no abnormal indicators and no early warnings were triggered. Users can use the evidence chain query module in the visual management console to enter the RPA task ID and export the full-chain audit report with one click. The report includes the task's full-process logs and hash verification information for compliance auditing.

[0052] 6. Regional rule updates and strategy optimization Every Monday, the regional compliance module automatically crawls the rule update documents of Douyin, Kuaishou, and TikTok. By parsing the rule text using the LLM algorithm, it discovers that TikTok Indonesia has added a new rule that "beauty products must not display skin comparison effects." This rule is immediately extracted, and the regional rule feature library is updated (the tolerance for vulgarity in Indonesia has been lowered, while the strictness of copyright has been increased). At the same time, the prompt constraints for content generation are dynamically adjusted (the new rule "skin comparison effects must not be displayed") and the risk control detection thresholds are adjusted (the sensitive word detection weight in the regional rule matching score is increased to 0.4) to ensure that subsequent published content complies with the latest compliance requirements.

[0053] Example 2: Suitable for multi-regional text and image publishing scenarios in news media This embodiment addresses the "multi-regional, multi-column, and highly timely" text and image publishing needs of news media. It incorporates the compliance requirements of platforms such as Toutiao, Tencent News, and Overseas Chinese Information Network to achieve automated content generation, regional compliance adaptation, scheduled publishing, and full-chain traceability. The specific implementation process is as follows: 1. System Deployment and Initialization Deploy the system described in this invention, configure an RPA resource pool (total capacity of 50 concurrent task nodes), and input 10 Toutiao accounts, 8 Tencent News accounts, and 5 overseas Chinese information network accounts (covering China, North America, and Europe). Through the regional compliance module, crawl the compliance rules of each platform and region, focusing on collecting politically sensitive words, copyright rules, and information publishing specifications (such as overseas information must not involve sensitive domestic political content), and construct a regional rule feature library, where the political sensitivity index is: China (high), North America (medium), and Europe (medium); the copyright strictness is: Tencent News (high), Toutiao (medium), and overseas Chinese information network (low).

[0054] 2. Content generation and semantic security optimization User input content production request: The topic is "Global Technology Industry Development Trends in 2026", the target regions are China, North America, and Europe, and the publishing channels are the corresponding accounts on the above three platforms. The generated initial Prompt is: "Summarize the global technology industry development trends in 2026, focusing on the analysis of artificial intelligence and new energy fields, with professional and objective language, suitable for information dissemination."

[0055] The content generation and semantic security layer calls the AIGC text model to generate three regionally adapted graphic and textual news articles (for China, North America, and Europe, respectively). Through the Prompt closed-loop optimization algorithm, it was found that the news in the Chinese region contained the statement "a certain overseas technology company monopolizes core technologies" (which has sensitive content), with a compliance rate of 58% (<60%). This triggered a 20% reduction in template weight and initiated manual review.

[0056] After manual review, the wording was adjusted to "A certain overseas technology company has strong competitiveness in core technology fields, and domestic companies are accelerating technological breakthroughs." The content was regenerated, and the compliance rate was calculated to be 96%. The reward function Reward = (1-0) × 0.7 + 89 / 100 × 0.3 = 0.7 + 0.267 = 0.967. At the same time, after every 500 iterations, the system automatically updates the template library, eliminating templates with lower compliance rates and adding optimized news-type Prompt templates.

[0057] 3. Risk control assessment and task scheduling The risk control and compliance engine calculates the R-score and C-score using a multi-model weighted fusion algorithm: BERT sensitive word matching score 90, GPT multimodal detection score (no illegal elements in text and images) 92, and regional rule matching score 94. The multi-model score = 90 × 0.3 + 92 × 0.4 + 94 × 0.3 = 27 + 36.8 + 28.2 = 92, corresponding to an R-score of 8 and a C-score of 92. The risk level is "low risk," and the regional compliance recommendation is "avoid sensitive and suggestive statements in information from the Chinese region, and pay attention to copyright citation guidelines in the North American and European regions."

[0058] The orchestration and dynamic control layer is based on R-score (8<30), allocates 80% of the RPA resource pool (40 concurrent nodes), and starts high-concurrency release; sets a timed release strategy (8 am in China, 8 pm in North America, and 3 pm in Europe, to adapt to the active time of users in each region), and limits the release frequency of Tencent News accounts (no more than 5 articles per hour) to avoid triggering platform traffic throttling.

[0059] 4. RPA Execution and Exception Handling The RPA task starts according to a scheduled strategy, automatically completing operations such as image and text editing, title optimization, topic addition, and submission. During the process, two Toutiao accounts encountered "page loading timeout" exceptions. The system automatically triggered the breakpoint resume mechanism, recording the current breakpoint (progress: "image and text editing completed, waiting to upload"). After a 30-second interval, it re-attempted to load the page. Upon successful loading, it resumed execution from the breakpoint, avoiding process interruptions that could cause publishing delays. At the same time, the monitoring module collected resource utilization in real time. If it found that some nodes had excessively high resource usage (over 90%), it automatically triggered resource scheduling, allocating some tasks from the high-utilization nodes to idle nodes to ensure stable process execution.

[0060] 5. Full-chain evidence and monitoring / early warning After publication, the end-to-end evidence chain module generates a three-level hash chain, linking the Prompt version ID, RPA task ID, risk control results (R-score=8 points, C-score=92 points), publication records (publication time, publishing account, content link), and user operation logs (manual review records, template optimization records), and storing them in a distributed ledger. Users can use the evidence chain query interface to input the hash of the published content to trace the entire process of the information, meeting compliance audit requirements.

[0061] The monitoring and early warning module collects core indicators in real time. During the release process, the success rate of prompt rewriting was 98%, the manual intervention rate was 2%, and the task timeout rate was 0%. All indicators did not exceed the preset thresholds, and no early warnings were triggered. Users can focus on the "R-score distribution" and "channel compliance rating changes" indicators through a custom monitoring panel to keep track of the release risk status in real time.

[0062] 6. Regional rule updates and strategy optimization The system automatically crawls rule updates from various platforms weekly. It was discovered that Tencent News had added a new rule: "News content must indicate the source of the cited information; otherwise, it will be considered infringement." The regional compliance module extracted this rule through rule parsing and adaptive algorithms, updating the regional rule feature library (Tencent News' copyright strictness has been increased to "extremely high"). Simultaneously, the Prompt constraints for content generation were dynamically adjusted (adding "information source indication"), and the multi-model weighted fusion algorithm of the risk control and compliance engine was adjusted, increasing the weight of the regional rule matching score to 0.4 to ensure that subsequently published information complies with Tencent News' latest compliance requirements.

[0063] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A risk control-driven RPA automated publishing system for media platforms, characterized in that, include: The content generation and semantic security layer integrates a Prompt closed-loop optimization algorithm, which is used to automatically rewrite the Prompt based on the type of risk control violation and iteratively generate compliant content through reinforcement learning. The risk control and compliance engine calculates R-score and C-score through a multi-model weighted fusion algorithm, and outputs risk level and regional compliance recommendations. The orchestration and dynamic control layer, based on the risk control-RPA linkage algorithm, adjusts the priority, concurrency, release channels and resource allocation strategies of RPA tasks in real time according to the R-score. When the R-score > 60, the task is automatically paused and a full-link rollback is triggered. The full-link evidence chain module, based on the SHA-256 hash chain and distributed ledger storage algorithm, links the Prompt version ID, RPA task ID, risk control results, release records and user operation logs into an immutable full-link evidence chain. The regional compliance module automatically crawls and extracts compliance rules for the target region / platform through rule parsing and adaptive algorithms, and dynamically adjusts content generation and review strategies. The monitoring and early warning module collects full-link metrics in real time. When the violation rate or rollback rate exceeds the threshold, it triggers automatic early warning and strategy optimization.

2. The risk control-driven RPA automated publishing system for media platforms according to claim 1, characterized in that, The Prompt closed-loop optimization algorithm includes a reinforcement learning iterative mechanism, specifically: Construct the reward function: Reward = (1 - probability of violation) × 0.7 + content quality score × 0.3, where the content quality score is calculated by weighting three dimensions: readability, relevance, and originality; When the compliance rate of the rewritten content is >95%, the weight of the corresponding template increases by 10%. When the compliance rate is less than 60%, the template weight is reduced by 20% and manual review is triggered; The template library is automatically updated every 500 iterations, adding high-quality templates and removing low-quality ones.

3. The risk control-driven RPA automated publishing system for media platforms according to claim 1, characterized in that, The multi-model weighted fusion algorithm of the risk control and compliance engine is specifically as follows: Multi-model score = BERT sensitive word matching score × 0.3 + GPT multimodal detection score × 0.4 + localization rule matching score × 0.3; Among them, the regional rule matching score is calculated based on the rule base built on the laws and regulations of the target region and the platform specifications, and is updated in real time.

4. The risk control-driven RPA automated publishing system for media platforms according to claim 1, characterized in that, The risk control-RPA linkage algorithm of the orchestration and dynamic adjustment layer also includes a dynamic resource scheduling strategy: When R-score < 30, allocate 80% of the RPA resource pool for high-concurrency deployments; When 30≤R-score<60, allocate 40% of the resource pool and restrict tasks to be executed only on channels with a compliance rating of ≥B; When R-score > 60, release all resources occupied by the task and trigger resource reflow, prioritizing allocation to low-risk tasks.

5. The risk control-driven RPA automated publishing system for media platforms according to claim 1, characterized in that, The rule parsing and adaptive algorithm of the regionalized compliance module includes: It automatically crawls the target platform's rule update documents every week, parses the rule text using the LLM algorithm, and extracts sensitive content keywords, compliance restrictions, and penalty rules. Construct a regional rule feature library, with feature dimensions including: political sensitivity index, copyright strictness, and tolerance for vulgar content; Based on the target area of ​​the RPA task, dynamically adjust the Prompt constraints and risk control detection thresholds for content generation.

6. The risk control-driven RPA automated publishing system for media platforms according to claim 1, characterized in that, The SHA-256 hash chain + distributed ledger storage algorithm of the full-link evidence chain module is as follows: A three-level hash chain is generated: content generation hash, risk control review hash, and release execution hash. Each level of hash is calculated based on the previous level of hash, forming an immutable association. It adopts a lightweight distributed ledger storage hash chain, which supports tracing the entire chain of information by any of the identifiers in the Prompt version ID, RPA task ID, and published content hash; The evidence chain query interface supports exporting compliance audit reports, which include the entire task process logs and hash verification information.

7. The risk control-driven RPA automated publishing system for media platforms according to claim 1, characterized in that, The monitoring and early warning module includes: Real-time collection of 12 core metrics: violation rate, rollback rate, success rate of Prompt rewriting, resource utilization rate, changes in channel compliance rating, frequency of regional rule updates, manual intervention rate, content production efficiency, R-score distribution, C-score distribution, task timeout rate, and user complaint rate. When any metric exceeds a preset threshold, a multi-channel alert is triggered, and strategy optimization suggestions are automatically generated. It supports customizable monitoring panels, allowing users to configure key metrics and alert thresholds according to their needs.

8. The risk control-driven RPA automated publishing system for media platforms according to claim 1, characterized in that, It also includes a visual management console with the following features: Real-time risk control dashboard: Displays R-score / C-score distribution, violation type statistics, and changes in channel compliance rating; RPA task orchestration view: Supports drag-and-drop adjustment of task flow, viewing task status and resource usage; Evidence chain query module: Enter the task ID or content hash to export a full-chain audit report with one click; Rule management interface: Manually edit compliance rules and upload custom Prompt rewrite templates.

9. A method for a risk control-driven RPA automated publishing system for media platforms, characterized in that... Includes the following steps: S1. Receive the content production request input by the user, extract the topic, target area, and publishing channel context information, and generate an initial Prompt; S2. Call the AIGC model to generate content, and calculate the R-score and C-score using a multi-model weighted fusion algorithm; S3. Adjust RPA task strategies based on the risk control-RPA linkage algorithm: If R-score < 30, allocate 80% of the resource pool to start a high-concurrency deployment; If 30 ≤ R-score < 60, allocate 40% of the resource pool and limit release to compliant channels; If the R-score is greater than 60, a Prompt closed-loop optimization is triggered. If the rewrite still violates the rules after 3 consecutive rewrites, a manual review is triggered. S4. After the release is completed, a three-level SHA-256 hash chain is generated and stored in the distributed ledger; S5. Real-time monitoring of core metrics; when a metric exceeds a threshold, an alert is triggered and optimization suggestions are generated. S6. Automatically crawl platform rules updates weekly, and update the regional rule feature library through rule parsing and adaptive algorithms.

10. The method for the risk control-driven RPA automated publishing system for media platforms according to claim 9, characterized in that, The Prompt closed-loop optimization in S3 includes: The risk control engine identifies the type of violation and retrieves the corresponding rewritten template from the template library. Generate an optimized Prompt, call AIGC to regenerate the content and calculate a new R-score; The template weights are updated based on the reinforcement learning reward function, and the subsequent rewriting logic is optimized.

Images