Cross-platform content marketing method and system based on multi-agent cooperation

Abstract

The application discloses a cross-platform content marketing method and system based on multi-agent cooperation, belongs to the technical field of multi-agent cooperation, and comprises the following steps: calculating fragmented data through association, mining business opportunities and extracting features, and constructing a real-time updated dynamic business opportunity knowledge graph.The application constructs a dynamic business opportunity knowledge graph, realizes multi-agent cooperative task scheduling, deep content generation, multi-modal real sense quantitative evaluation, real-time risk control intervention and evolution optimization of the whole process closed loop, significantly improves the authenticity of marketing content and user resonance, compresses the whole link cycle from content generation to release from days to real time, shortens the risk response time from hours to milliseconds, improves the marketing ROI, improves the content real sense index, reduces the marketing risk rate, realizes the whole link intelligentization from business opportunity mining to strategy optimization, and improves the user conversion efficiency.

Description

Technical Field

[0001] This invention belongs to the field of multi-agent collaboration technology, specifically referring to a cross-platform content marketing method and system based on multi-agent collaboration. Background Technology

[0002] With the rapid development of mobile internet and social media, content marketing has become a core means for enterprises to acquire traffic, build brand awareness, and promote conversion. Cross-platform operation requires marketing entities to quickly produce and distribute massive amounts of high-quality content based on the algorithm mechanisms, user profiles, and content ecosystems of different platforms. Faced with increasingly fierce market competition and the fragmentation of user attention, traditional content marketing models and existing automated marketing tools have gradually revealed many limitations and are unable to meet the current market's stringent requirements for high authenticity, strong persuasiveness, and real-time risk control.

[0003] However, multi-agent collaborative cross-platform content marketing methods still have certain shortcomings. Existing technologies struggle to extract commercially valuable relevance signals from massive amounts of fragmented information; content homogenization exists, with copywriting produced by large models exhibiting a strong AI feel that fails to match the authentic, grassroots, and lifestyle-oriented community tone of Xiaohongshu; processes are non-standardized, making it difficult to accumulate marketing experience, with creation efficiency and quality heavily reliant on human intervention, resulting in low user resonance, inability to accurately trigger user cognitive actions, and insufficient persuasiveness; in the authenticity assessment stage, existing technologies rely on subjective judgment or simple rules, leading to incomplete and unquantifiable assessments, long optimization cycles, and difficulty in real-time improvement of content quality, failing to meet users' demands for authenticity; in the risk monitoring stage, risk identification is lagging, intervention is passive, and the marketing process is neglected after release, making it difficult to control risks in a timely manner, leading to marketing activities easily spiraling out of control; and in the strategy optimization stage, a cross-platform content marketing method and system based on multi-agent collaboration is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a cross-platform content marketing method and system based on multi-agent collaboration to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a cross-platform content marketing method based on multi-agent collaboration, comprising the following steps:

[0006] S1. By associating fragmented data, perform calculations to mine business opportunities and extract features, and build a dynamic business opportunity knowledge graph that is updated in real time;

[0007] S2. Based on the constructed dynamic business opportunity knowledge graph, dynamic task release and negotiation scheduling among various functional intelligent agents are carried out through the collaborative central intelligent agent;

[0008] S3. Based on the tasks assigned by the collaboration center and fragments in the business opportunity knowledge graph, the intention understanding sub-agent analyzes the user's psychology and generates deeply persuasive content.

[0009] S4. Based on the generated content, a multi-modal evaluation agent performs multi-dimensional quantitative scoring and feeds the realism evaluation results back to the creation end in real time.

[0010] S5. Based on the optimized content and preset strategies, the system integrates a real-time risk control intelligent agent to perform release execution and full-process risk monitoring and intervention.

[0011] S6. Based on the marketing performance data collected across the entire value chain, optimize and iterate through an evolutionary learning agent.

[0012] Preferably, in step S1, the system continuously collects fragmented data from the platform via a crawler interface and the platform's open API. The collected raw data is cleaned, and natural language processing technology is used for key entity recognition and preliminary relationship labeling. Unstructured text is converted into semi-structured data. Through a pre-trained model and custom rules, commercial feature vectors are extracted from the pre-processed data. These feature vectors are then input into an association rule mining and preset graph neural network model to calculate the implicit association strength between different feature entities. This process is as follows:

[0013] ,

[0014] In the formula, Representing characteristic entities and The strength of the implicit association, Represents a node The feature vector updated by L layers of GNN, Represents a node The feature vector updated by L layers of GNN, express Euclidean norm, express The Euclidean norm, where L represents the number of layers in the GNN. for , This represents the feature vector of node v at layer l. Let W represent the set of neighboring nodes of node v, and let W represent the weight matrix. This represents the activation function, and the final feature vector of node v after passing through L layers of a GNN. .

[0015] The association rules are as follows:

[0016] ,

[0017] In the formula, Represents cosine similarity. Representing the eigenvector and dot product, and Representing the eigenvector and The model.

[0018] The discovered strong correlation patterns are organized into triples of entity-relationship-attribute and stored in a graph database to form a dynamic business opportunity knowledge graph. Entities represent business elements, relationships represent the business logic connections between them, and attributes record popularity, confidence level, and timestamps. The system establishes real-time data stream monitoring, processes new data streams, and updates the dynamic business opportunity knowledge graph incrementally. Based on user interaction data, such as new notes added to favorites and changes in sentiment in comments, the weights of existing nodes and relationships in the dynamic business opportunity knowledge graph are dynamically adjusted.

[0019] Preferably, in step S2, the collaborative central agent decomposes marketing objectives into atomic-level subtasks based on the real-time attributes of nodes in the dynamic business opportunity knowledge graph. Marketing objectives include, for example, increasing the interaction rate of hyaluronic acid essence product notes. These atomic-level subtasks are described using structured task descriptions, including task type, dependencies, expected benefit indicators, and timeliness requirements. These descriptions are then published to the system's internal dynamic task market. The task market maintains a task queue and agent state pool in real time, including the load rate, resource availability, and skill vectors of each functional agent. The system maintains the agent state pool and updates the state of each functional agent in real time, including load rate, resource availability, and AgentSkill library skill vectors.

[0020] Preferably, in step S2, each functional agent, based on its own predefined AgentSkill library, such as skill vectors (e.g., the pain point analysis ability and emotional expression ability of the content creation agent), and real-time status (e.g., current task load rate and GPU resource utilization rate), dynamically bids for sub-tasks in the task market. The benefit evaluation of the collaboration center is quantitatively scored, thus achieving the following:

[0021] ,

[0022] In the formula, E represents the total benefit score, and M represents the dynamic skill matching degree:

[0023] ,

[0024] in, This represents the task skill requirement weight, indicating the intensity of the task's demand for the i-th skill, with a value range of [0, 1]. This represents the agent's skill vector value, which is the agent's ability value in the i-th skill, and its value ranges from [0, 1]. This represents the urgency factor of the i-th skill in the task, reflecting the urgency of the skill in the current marketing scenario, with a value range of [0, 1], where n represents the total number of skill dimensions. This represents the skill matching weight coefficient, with a value range of [0, 1], and L represents the load balancing degree.

[0025] ,

[0026] Where c represents the current load rate of the agent, with a value in the range [0, 1], and a represents the average load rate of the system, with a value in the range [0, 1]. This represents the load balancing weighting coefficient. ;

[0027] Based on the evaluation results, all bids are ranked, and the solution with the highest overall benefit score is selected. A task contract is then signed with the bidding agent, and the collaboration center continuously monitors the process during task execution.

[0028] Changes in node weights of the dynamic business opportunity knowledge graph;

[0029] External environmental events;

[0030] If the trigger condition is detected, renegotiation will be performed immediately:

[0031] (a) Pause the current task flow;

[0032] (b) Based on the updated knowledge graph, task priorities are re-decomposed, such as increasing the priority of high-weight node tasks by 2 levels;

[0033] (c) Publish the optimized task announcement in the task marketplace;

[0034] (d) Complete the re-bidding and allocation within the preset time;

[0035] Upon completion of the task, the central collaborative agent verifies the outputs of each agent and integrates them into a structured collaborative output package, which includes:

[0036] Task execution log, including agent bidding scores and renegotiation trigger records;

[0037] Optimized content materials, including text and image combinations and sentiment analysis reports;

[0038] Degree of achievement of performance indicators.

[0039] Preferably, in step S3, the structured task and related dynamic business opportunity knowledge graph fragments assigned by the collaboration center are received, and based on a pre-trained intent recognition model, the fragments are deeply analyzed to calculate the intent confidence level, which is implemented as follows:

[0040] ,

[0041] In the formula, This indicates the confidence level of the intent, with a value range of [0, 1]. The weight of user pain points is represented by the node attributes of the business opportunity knowledge graph, reflecting the intensity of the pain points. 'e' represents the sentiment score. It represents minute constants, identifies core user intent types, and outputs an intent parsing report.

[0042] Preferably, in step S3, the intent understanding sub-agent constructs structured content instructions based on the intent parsing report, including persuasive logic (e.g., resolving anxiety, or providing value recognition), emotional tone intensity (e.g., resonant sharing, professional sharing), and the expected user cognitive actions; the emotional tone is achieved as follows: d represents the intensity of the emotional tone, the magnitude of the emotional shift achieved through the content. This represents the target sentiment score. The content creation agent uses structured content instructions, combined with a finely tuned underlying large language model and visual generation model. Simultaneously, the internal prompt word engineering dynamically constructs optimal and highly specific prompt words based on the instructions, driving the model to create content. The generated initial draft of the content will first undergo a rapid self-check within the creation agent. The self-check criteria include: alignment with the structured content instructions, accuracy of basic facts, and whether it conforms to the preset baseline of amateur characteristics, such as sentence complexity and colloquialism. If it does not meet the standards, feedback is provided for parameter adjustment and regeneration.

[0043] Preferably, in step S4, the generated content is acquired, semantic features are extracted from the text, and based on a pre-trained text analysis model, the degree of colloquialism, emotional fluctuation patterns, and density of lifelike expressions are identified. Visual features are extracted from the image, and based on a convolutional neural network, the degree of lifelikeness of the scene, the naturalness of the composition, and the realism of visual elements are identified. By analyzing the text features, a text dimension realism score is calculated, thus achieving the following:

[0044] ,

[0045] In the formula, This represents the text dimension's realism score, with a value range of [0, 1]. Indicates colloquial weighting, Indicates the degree of colloquialism in the text, with a value range of [0, 1]. Indicates the weight of emotional fluctuations. This indicates the intensity of the text's emotional fluctuation, with a value range of [0, 1]. Represents a small constant;

[0046] By identifying scene elements and the unintentionality of composition in an image, a visual dimension realism score is generated, achieving the following:

[0047] ,

[0048] In the formula, This represents the visual realism score, with a value range of [0, 1]. Indicates the weight of visual naturalness. This represents the intensity of visual naturalness, with a value range of [0, 1]. Indicates scene weight, Scene coverage, with a value range of [0, 1]. This represents a small constant.

[0049] Preferably, in step S4, the comprehensive realism index is output by dynamically weighting and integrating the text-dimensional realism score and the visual-dimensional score, as follows:

[0050] ,

[0051] In the formula, To convey a sense of overall realism, This represents the consistency score between the text and images, with a value range of [0, 1]. Indicates the dynamic weight of the dimension;

[0052] The overall realism index is returned to the content generation process in real time;

[0053] If the index falls below the system's preset dynamic threshold, the content optimization process will be triggered.

[0054] If the index meets the requirements, the content will directly enter the publishing process.

[0055] Preferably, in step S5, after obtaining the optimized content and preset release strategy, before the official release, the real-time risk control intelligence agent initiates a pre-release review, calls the compliance library synchronized with the platform rules, performs a final round of compliance and security verification on the content, and checks for conflicts in the release strategy. After passing the pre-release review, the system releases the content to the target platform according to the strategy. Simultaneously, the system embeds full-link monitoring points into the content to track core indicators after release. The real-time risk control intelligence agent starts synchronously and begins listening to the data stream. After release, the real-time risk control intelligence agent enters a continuous monitoring state, collects the content's performance data in real time, and performs correlation analysis in conjunction with the current state of the dynamic business opportunity knowledge graph. Through a preset risk model, it dynamically calculates the real-time risk score of the content. The implementation is as follows:

[0056] ,

[0057] In the formula, This represents the weight of the i-th risk factor. This represents the quantified value of risk factors.

[0058] The assessment dimensions include, but are not limited to: interaction rate health, sudden changes in comment sentiment, and sudden negative related topic popularity. When the risk score exceeds a certain threshold, the risk control intelligence agent automatically matches and executes the corresponding SOP (Standard Operating Procedure) from the preset intervention strategy library according to the risk level and type.

[0059] Preferably, in step S6, the evolutionary learning agent initiates a data pipeline to acquire all relevant data generated throughout the entire process from S1 to S5, including: a snapshot of the dynamic business opportunity knowledge graph in S1, task allocation records and agent performance data in S2, structured content instructions in S3, a multi-dimensional realism assessment report and score in S4, content publishing data in S5, real-time risk control logs, and final marketing effect data obtained from the platform API, such as interaction rate, number of new followers, inclusion status, and conversion rate. The evolutionary learning agent performs attribution analysis on each marketing case, quantifying the contribution of different factors to the final effect through statistical analysis, such as A / B testing comparison and machine learning models. Based on the results of the attribution analysis, the evolutionary learning agent reverse-optimizes the various core knowledge bases and strategy bases in the system.

[0060] Optimize the dynamic business opportunity knowledge graph: Valid business opportunity nodes will have their weight and confidence increased; invalid or outdated nodes will have their weight reduced or archived. At the same time, new connections revealed in success cases will be added to the graph as new edges.

[0061] Extracting and generating new AgentSkills: Abstracting reusable operation patterns from successful task execution records and formalizing them into new or improved AgentSkills, registering them in the skill library of the corresponding functional intelligent agent for use during task negotiation and scheduling;

[0062] Evolutionary Marketing SOP: If a pain point with a weight in the knowledge graph that has reached a preset threshold is found, and the instruction to resonate with the audience through an emotional tone is consistently successful, then this combination can be solidified into a new SOP.

[0063] If it is found that a certain type of risk control intervention SOP can always effectively curb the spread of negative effects, then the priority and trigger threshold of the SOP should be increased.

[0064] The evolutionary learning agent continuously fine-tunes the key models in the system online through accumulated high-quality case-effect data. The evolutionary learning agent performs A / B testing on these updates in a simulation environment or a low-traffic experimental environment. After verification, the system configures policies through the web interface and performs mobile monitoring and one-click distribution through the mini-program interface.

[0065] Preferably, it includes a dynamic business opportunity knowledge graph construction module, a multi-agent dynamic collaboration scheduling module, an intent-driven content generation module, a multimodal realism assessment module, a full-link real-time risk control execution module, and a strategy evolution optimization module.

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] 1. This invention significantly improves the authenticity and user resonance of marketing content by constructing a dynamic business opportunity knowledge graph, realizing multi-agent collaborative task scheduling, deep content generation, multimodal realism quantitative evaluation, real-time risk control intervention, and evolution optimization in a closed-loop process. It compresses the entire chain cycle from content generation to release from days to real time, shortens the risk response time from hours to milliseconds, improves marketing ROI, enhances the content authenticity index, and reduces marketing risk rate. It achieves full-chain intelligence from business opportunity mining to strategy optimization and improves user conversion efficiency.

[0068] 2. This invention, through the intention understanding sub-agent's deep analysis of user psychology and the construction of structured content instructions, makes content creation more aligned with users' real needs and emotional changes, realizing the transformation of content from passive generation to active persuasion, significantly improving the persuasiveness of content and user participation, enabling content to accurately trigger user cognitive actions and improve conversion rates.

[0069] 3. This invention uses a multimodal evaluation agent to perform multi-dimensional quantitative scoring and real-time feedback mechanisms, which realizes objective evaluation and instant optimization of the authenticity of content, enabling the quality of content to be improved in real time, enhancing users' recognition of the authenticity of the content, and ensuring that the generated content has high life-like characteristics and strong user resonance.

[0070] 4. This invention achieves proactive risk identification and precise intervention through real-time risk control intelligence, enabling the marketing process to respond to risk changes in real time, significantly improving marketing security and success rate, ensuring that content publication complies with platform specifications and user expectations, and reducing marketing risk rate. Attached Figure Description

[0071] Figure 1 The following is the operational flow of the cross-platform content marketing method based on multi-agent collaboration of this invention. Figure 1 ;

[0072] Figure 2 The following is the operational flow of the cross-platform content marketing method based on multi-agent collaboration of this invention. Figure 2 ;

[0073] Figure 3 The following is the operational flow of the cross-platform content marketing method based on multi-agent collaboration of this invention. Figure 3 ;

[0074] Figure 4 The following is the operational flow of the cross-platform content marketing method based on multi-agent collaboration of this invention. Figure 4 ;

[0075] Figure 5 The following is the operational flow of the cross-platform content marketing method based on multi-agent collaboration of this invention. Figure 5 ;

[0076] Figure 6 The following is the operational flow of the cross-platform content marketing method based on multi-agent collaboration of this invention. Figure 6 ;

[0077] Figure 7 This is a schematic diagram of the cross-platform content marketing system based on multi-agent collaboration of the present invention. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] Example

[0080] Please see Figures 1-7 As shown, the present invention provides a technical solution comprising the following steps:

[0081] S1. By associating fragmented data, perform calculations to mine business opportunities and extract features, and build a dynamic business opportunity knowledge graph that is updated in real time;

[0082] S2. Based on the constructed dynamic business opportunity knowledge graph, dynamic task release and negotiation scheduling among various functional intelligent agents are carried out through the collaborative central intelligent agent;

[0083] S3. Based on the tasks assigned by the collaboration center and fragments in the business opportunity knowledge graph, the intention understanding sub-agent analyzes the user's psychology and generates deeply persuasive content.

[0084] S4. Based on the generated content, a multi-modal evaluation agent performs multi-dimensional quantitative scoring and feeds the realism evaluation results back to the creation end in real time.

[0085] S5. Based on the optimized content and preset strategies, the system integrates a real-time risk control intelligent agent to perform release execution and full-process risk monitoring and intervention.

[0086] S6. Based on the marketing performance data collected across the entire value chain, optimize and iterate through an evolutionary learning agent.

[0087] In this embodiment, in step S1, the system continuously collects fragmented data from the platform via a web crawler interface and the platform's open API. This data includes note text, comments, user profiles, search hot words, topic tags, and competitor account dynamics. The collected raw data is cleaned, including noise reduction, deduplication, and standardization. Natural language processing technology is used for key entity recognition, such as product, ingredient, scene, sentiment words, and preliminary relationship labeling. Unstructured text is converted into semi-structured data. Through a pre-trained model and custom rules, commercial feature vectors are extracted from the pre-processed data, including but not limited to: demand intensity, sentiment tendency, topic popularity, user group relevance, and content scarcity. The feature vectors are input into association rule mining and a preset graph neural network model to calculate the implicit association strength between different feature entities. This is achieved as follows:

[0088] ,

[0089] In the formula, Representing characteristic entities and The strength of the implicit association, Represents a node The feature vector updated by L layers of GNN, Represents a node The feature vector updated by L layers of GNN, express Euclidean norm, express The Euclidean norm, where L represents the number of layers in the GNN. for , This represents the feature vector of node v at layer l. Let W represent the set of neighboring nodes of node v, and let W represent the weight matrix. This represents the activation function, and the final feature vector of node v after passing through L layers of a GNN. .

[0090] The association rules are as follows:

[0091] ,

[0092] In the formula, Represents cosine similarity. Representing the eigenvector and dot product, and Representing the eigenvector and The model.

[0093] The discovered strong correlation patterns are organized into triples of entity-relationship-attribute and stored in a graph database to form a dynamic business opportunity knowledge graph. Entities represent business elements, such as user pain points, products, and ingredients; relationships represent the business logic connections between them, such as triggering, applicable, and matching; and attributes record popularity, confidence level, and timestamps. The system establishes real-time data stream monitoring, processes new data streams, and updates the dynamic business opportunity knowledge graph incrementally. Based on user interaction data, such as new notes added to favorites and changes in sentiment in comments, the weights of existing nodes and relationships in the dynamic business opportunity knowledge graph are dynamically adjusted.

[0094] In this embodiment, in step S2, the collaborative central agent decomposes marketing objectives into atomic-level sub-tasks based on the real-time attributes of nodes in the dynamic business opportunity knowledge graph. Marketing objectives include, for example, increasing the interaction rate of hyaluronic acid essence product notes. Atomic-level sub-tasks include, for example, generating content addressing pain points related to dry and sensitive skin with enlarged pores, and designing lifestyle-oriented image and text combinations. The sub-tasks are described in a structured task description, including task type, dependencies, expected benefit indicators, and timeliness requirements, and are published to the dynamic task market within the system. The task market maintains the task queue and agent state pool in real time, including the load rate, resource availability, and skill vectors of each functional agent. The system maintains the agent state pool and updates the state of each functional agent in real time, including the load rate, resource availability, and AgentSkill library skill vectors.

[0095] In this embodiment, in S2, each functional intelligent agent, such as the content creation intelligent agent, the review intelligent agent, and the publishing intelligent agent, dynamically bids for sub-tasks in the task market based on its own predefined AgentSkill library, such as skill vectors, such as the pain point analysis ability and emotional expression ability of the content creation intelligent agent, and real-time status, such as the current task load rate and GPU resource utilization rate. The benefit evaluation of the collaboration center is carried out by quantitative scoring, which is implemented as follows:

[0096] ,

[0097] In the formula, E represents the total benefit score, and M represents the dynamic skill matching degree:

[0098] ,

[0099] in, This represents the task skill requirement weight, indicating the intensity of the task's demand for the i-th skill, with a value range of [0, 1]. This represents the agent's skill vector value, which is the agent's ability value in the i-th skill, and its value ranges from [0, 1]. This represents the urgency factor of the i-th skill in the task, reflecting the urgency of the skill in the current marketing scenario, with a value range of [0, 1], where n represents the total number of skill dimensions. This represents the skill matching weight coefficient, with a value range of [0, 1], and L represents the load balancing degree.

[0100] ,

[0101] Where c represents the current load rate of the agent, with a value in the range [0, 1], and a represents the average load rate of the system, with a value in the range [0, 1]. This represents the load balancing weighting coefficient. ;

[0102] Based on the evaluation results, all bids are ranked, and the solution with the highest overall benefit score is selected. A task contract is then signed with the bidding agent, and the collaboration center continuously monitors the process during task execution.

[0103] Changes in node weights of the dynamic business opportunity knowledge graph;

[0104] External environmental events, such as sudden changes in user comment sentiment scores exceeding 0.2;

[0105] If the trigger condition is detected, renegotiation will be performed immediately:

[0106] (a) Pause the current task flow;

[0107] (b) Based on the updated knowledge graph, task priorities are re-decomposed, such as increasing the priority of high-weight node tasks by 2 levels;

[0108] (c) Publish the optimized task announcement in the task marketplace;

[0109] (d) Complete the re-bidding and allocation within the preset time;

[0110] Upon completion of the task, the central collaborative agent verifies the outputs of each agent, such as whether the notes delivered by the content creation module meet the emotional tone requirements, and integrates them into a structured collaborative output package, which includes:

[0111] Task execution log, including agent bidding scores and renegotiation trigger records;

[0112] Optimized content materials, including text and image combinations and sentiment analysis reports;

[0113] Degree of achievement of performance indicators.

[0114] In this embodiment, step S3 involves receiving structured tasks and related dynamic business opportunity knowledge graph fragments assigned by the collaboration center, performing deep analysis on the fragments based on a pre-trained intent recognition model, and calculating intent confidence. This is achieved as follows:

[0115] ,

[0116] In the formula, This indicates the confidence level of the intent, with a value range of [0, 1]. The weight of user pain points is represented by the node attributes of the business opportunity knowledge graph, reflecting the intensity of the pain points. 'e' represents the sentiment score. Representing minute constants, identifying core user intent types, such as seeking solutions with a confidence level greater than or equal to... The confidence level of the desire for identity recognition is greater than or equal to Output an intent analysis report.

[0117] In this embodiment, in step S3, the intent understanding sub-agent constructs structured content instructions based on the intent parsing report, including persuasive logic (e.g., resolving anxiety or providing value recognition), emotional tone intensity (e.g., resonant sharing, professional sharing), and the expected user cognitive actions (e.g., from ignorance to knowledge, from doubt to trust); the emotional tone is implemented as follows: d represents the intensity of the emotional tone, the magnitude of the emotional shift achieved through the content. This represents the target sentiment score. The content creation agent uses structured content instructions, combined with a finely tuned underlying large language model and visual generation model. Simultaneously, the internal prompt word engineering dynamically constructs optimal and highly specific prompt words based on the instructions, driving the model to create content. The generated initial draft of the content will first undergo a rapid self-check within the creation agent. The self-check criteria include: alignment with the structured content instructions, accuracy of basic facts, and whether it conforms to the preset baseline of amateur characteristics, such as sentence complexity and colloquialism. If it does not meet the standards, feedback is provided for parameter adjustment and regeneration.

[0118] In this embodiment, in step S4, the generated content is acquired, semantic features are extracted from the text, and based on a pre-trained text analysis model, the degree of colloquialism, emotional fluctuation patterns, and density of lifelike expressions are identified. Visual features are extracted from the image, and based on a convolutional neural network, the degree of lifelikeness of the scene, the naturalness of the composition, and the realism of visual elements are identified. By analyzing the text features, a text dimension realism score is calculated, which is implemented as follows:

[0119] ,

[0120] In the formula, This represents the text dimension's realism score, with a value range of [0, 1]. Indicates colloquial weighting, Indicates the degree of colloquialism in the text, with a value range of [0, 1]. Indicates the weight of emotional fluctuations. This indicates the intensity of the text's emotional fluctuation, with a value range of [0, 1]. Represents a small constant;

[0121] By identifying scene elements and the unintentionality of composition in an image, a visual dimension realism score is generated, achieving the following:

[0122] ,

[0123] In the formula, This represents the visual realism score, with a value range of [0, 1]. Indicates the weight of visual naturalness. This represents the intensity of visual naturalness, with a value range of [0, 1]. Indicates scene weight, Scene coverage, with a value range of [0, 1]. This represents a small constant.

[0124] In this embodiment, in step S4, a comprehensive realism index is output by dynamically weighting and integrating the text dimension realism score and the visual dimension score, which is implemented as follows:

[0125] ,

[0126] In the formula, To convey a sense of overall realism, This represents the consistency score between the text and images, with a value range of [0, 1]. Indicates the dynamic weight of the dimension;

[0127] The overall realism index is returned to the content generation process in real time;

[0128] If the index falls below the system's preset dynamic threshold, such as the historical average index based on similar content, then the content optimization process is triggered:

[0129] (a) Evaluation results, such as insufficient colloquialism in the text and lack of lifelike scenarios, are fed back to the intention-understanding sub-agent in the form of structured instructions;

[0130] (b) The intention-understanding sub-agent adjusts the content generation instructions based on feedback, such as enhancing the requirements for colloquial language and supplementing the constraints of everyday scenarios;

[0131] (c) The content generation module regenerates the content and enters a new round of evaluation.

[0132] If the index meets the requirements, the content will directly enter the publishing process.

[0133] In this embodiment, in step S5, the optimized content and preset publishing strategy are obtained. Before the official release, the real-time risk control intelligence initiates a pre-release review, calling a compliance library synchronized with platform rules to perform a final round of compliance and security verification on the content, and checks for conflicts in the publishing strategy. After passing the pre-release review, the system publishes the content to the target platform according to the strategy. Simultaneously, the system embeds end-to-end monitoring points into the content to track core metrics after release, such as exposure, click-through rate, interaction data, and user comments. The real-time risk control intelligence starts simultaneously and begins listening to the data stream. After release, the real-time risk control intelligence enters a continuous monitoring state, collecting the content's performance data in real time and performing correlation analysis based on the current state of the dynamic business opportunity knowledge graph. Through a preset risk model, the real-time risk score of the content is dynamically calculated. The implementation is as follows:

[0134] ,

[0135] In the formula, This represents the weight of the i-th risk factor. This represents the quantified value of risk factors.

[0136] The assessment dimensions include, but are not limited to: interaction rate health, sudden changes in comment sentiment, and sudden negative related topic popularity. When the risk score exceeds a certain threshold, the risk control intelligence agent automatically matches and executes the corresponding SOP from the preset intervention strategy library according to the risk level and type.

[0137] In this embodiment, in step S6, the evolutionary learning agent initiates a data pipeline to acquire all relevant data generated throughout the entire process from S1 to S5, including: a snapshot of the dynamic business opportunity knowledge graph in S1, task allocation records and agent performance data in S2, structured content instructions in S3, a multi-dimensional realism assessment report and score in S4, content publishing data in S5, real-time risk control logs, and final marketing effect data obtained from the platform API, such as interaction rate, number of new followers, inclusion status, and conversion rate. The evolutionary learning agent performs attribution analysis on each marketing case, quantifying the contribution of different factors to the final effect through statistical analysis, such as A / B testing comparison and machine learning models. Based on the results of the attribution analysis, the evolutionary learning agent reverse-optimizes the various core knowledge bases and strategy bases in the system.

[0138] Optimize the dynamic business opportunity knowledge graph: Valid business opportunity nodes will have their weight and confidence increased; invalid or outdated nodes will have their weight reduced or archived. At the same time, new connections revealed in success cases will be added to the graph as new edges.

[0139] Extracting and generating new AgentSkills: Abstracting reusable operation patterns from successful task execution records and formalizing them into new or improved AgentSkills, registering them in the skill library of the corresponding functional intelligent agent for use during task negotiation and scheduling;

[0140] Evolutionary Marketing SOP: If a pain point with a weight in the knowledge graph that has reached a preset threshold is found, and the instruction to resonate with the audience through an emotional tone is consistently successful, then this combination can be solidified into a new SOP.

[0141] If it is found that a certain type of risk control intervention SOP can always effectively curb the spread of negative effects, then the priority and trigger threshold of the SOP should be increased.

[0142] The evolutionary learning agent continuously fine-tunes the key models in the system online through accumulated high-quality case-effect data. The evolutionary learning agent performs A / B testing on these updates in a simulation environment or a low-traffic experimental environment. After verification, the system configures policies through the web interface and performs mobile monitoring and one-click distribution through the mini-program interface.

[0143] This embodiment includes a dynamic business opportunity knowledge graph construction module, a multi-agent dynamic collaboration scheduling module, an intent-driven content generation module, a multimodal realism assessment module, a full-link real-time risk control execution module, and a strategy evolution optimization module.

[0144] In this embodiment, the collaborative central intelligent agent includes a task decomposition engine (which breaks down marketing objectives into atomic-level sub-tasks), a dynamic bidding evaluator (which calculates the total benefit score), and a renegotiation trigger (which performs task rescheduling based on changes in the business opportunity graph).

[0145] Global Coordination: Receives the dynamic business opportunity knowledge graph generated by S1 and decomposes marketing objectives into executable atomic-level subtasks;

[0146] Task scheduling: Maintain the task market and agent state pool, and execute dynamic bidding ordering and contract signing;

[0147] Process monitoring: Continuously monitor changes in the weights of the business opportunity map and external environmental events to trigger task renegotiation.

[0148] The intent understanding sub-agent includes a user mental parser (calculating intent confidence), a structured instruction generator (outputting persuasion logic, emotional tone, and cognitive action), and a cue word engineer (dynamically constructing highly specific cue words).

[0149] In-depth demand analysis: Based on the structured tasks and business opportunity knowledge graph fragments allocated by S2, analyze user psychology;

[0150] Content instruction construction: Generate structured instructions containing persuasive logic (resolving anxiety), emotional tone (empathetic expression), and cognitive action (from doubt to trust). The intention is to understand the prompts and structured instructions output by the sub-agent and pass them to the content creation agent, which drives the fine-tuned model to complete content generation and perform a first draft self-check.

[0151] Command control: Output intent parsing report.

[0152] The multimodal evaluation agent includes a text feature extractor (analyzing colloquialisms and emotional fluctuations), a visual feature extractor (analyzing scene realism and composition naturalness), and a realism integrator (dynamically weighting and outputting a comprehensive realism index).

[0153] Multi-dimensional scoring: The content generated by S3 is quantitatively scored in terms of both textual and visual dimensions;

[0154] Real-time feedback: The overall realism index is returned to the S3 content generation stage in real time, triggering an optimization process;

[0155] Quality inspection: Ensure the content meets the standards, and proceed to the S5 release process once the overall realism is greater than or equal to the dynamic threshold.

[0156] The real-time risk control intelligent agent includes a compliance pre-screener (which calls the platform rule base to verify content), a dynamic risk calculation engine (which calculates risk scores in real time), and an SOP matching executor (which automatically matches intervention strategies).

[0157] Pre-release verification: Before S5 is launched, the content undergoes a final review for compliance and security.

[0158] Full-process monitoring: Real-time collection of interaction rate, sentiment polarity, and negative association popularity data after publication, and dynamic calculation of risk score;

[0159] Proactive intervention: When the risk score exceeds the threshold, an SOP is automatically matched and executed.

[0160] Evolutionary learning agents include a full-link data integrator (aggregating data from the entire process of S1 to S5), a policy optimizer (optimizing policy parameters through genetic algorithms), and a knowledge base evolver (updating the business opportunity graph and AgentSkill library).

[0161] Data closed loop: Collect all data generated from S1 to S5, including knowledge graph snapshots, task logs, realism scores, and risk control logs;

[0162] Strategy optimization: Optimize content instruction weights and risk control threshold coefficients based on evolutionary learning;

[0163] Knowledge base update: Transform effective strategies into new AgentSkills.

[0164] Working principle: The system continuously collects fragmented data through web crawlers and platform APIs. After cleaning, it uses natural language processing technology to identify key entities and label relationships, transforming unstructured text into semi-structured data. Business feature vectors are extracted and input into association rule mining and graph neural network models to calculate the strength of implicit associations between feature entities. The data is then organized into entity-relationship-attribute triples and stored in a graph database to form a dynamic business opportunity knowledge graph. The system monitors the data stream in real time, incrementally updates the graph content, and dynamically adjusts the weights of nodes and relationships based on user interaction data.

[0165] The collaborative central agent decomposes marketing objectives into atomic-level subtasks based on the real-time attributes of nodes in the dynamic business opportunity knowledge graph, and publishes them to the dynamic task marketplace with structured descriptions. Each functional agent dynamically bids for subtasks based on its own skill vectors and real-time status. The collaborative central agent calculates the overall benefit score, ranks the subtasks, and selects the optimal solution. During task execution, it continuously monitors changes in knowledge graph weights and external environmental events, triggering renegotiation to optimize task priorities. After task completion, it integrates collaborative output packages. The intent understanding sub-agent receives structured tasks and knowledge graph fragments, deeply analyzes user psychology, and calculates intent confidence. Based on the analysis report, it constructs structured content instructions containing persuasive logic, emotional tone intensity, and user cognitive actions. The prompt word engineering dynamically generates highly specific prompt words to drive model creation. The initial content draft undergoes rapid self-checking within the creation agent to ensure alignment with instructions, factual accuracy, and human-like characteristics; if it fails to meet the standards, feedback is provided for adjustment and regeneration. The multimodal evaluation agent acquires the generated content, extracts semantic features from the text, and identifies the degree of colloquialism and emotional fluctuations. The system extracts visual features from images to identify the lifelikeness of the scene and the naturalness of the composition; it calculates realism scores for text and visual dimensions, and outputs a comprehensive realism index through dynamic weighting and integration; the index is fed back to the creation end in real time, triggering a content optimization process if it falls below a dynamic threshold, and entering the release process if it meets the standard; the real-time risk control agent calls the compliance library to verify the content and strategy before release, and releases the content and embeds full-link monitoring points after the pre-approval; the risk control agent continuously monitors the data stream, collects performance data and combines it with knowledge graph state correlation analysis to dynamically calculate risk scores; when the risk score exceeds the threshold, it automatically matches and executes intervention SOPs to achieve proactive monitoring and precise intervention of risks throughout the marketing process; the evolutionary learning agent collects full-process data, performs attribution analysis to quantify the contribution of factors; based on the analysis results, it optimizes the weight of the dynamic business opportunity knowledge graph, extracts new Agent Skills, and evolves marketing SOPs; after the system is verified and updated in the simulation environment, it achieves continuous evolution and closed-loop optimization of strategies through web configuration and mini-program monitoring and distribution.

[0166] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

[0167] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A cross-platform content marketing method based on multi-agent collaboration, characterized in that, Includes the following steps: S1. By associating fragmented data, perform calculations to mine business opportunities and extract features, and build a dynamic business opportunity knowledge graph that is updated in real time; S2. Based on the constructed dynamic business opportunity knowledge graph, dynamic task release and negotiation scheduling among various functional intelligent agents are carried out through the collaborative central intelligent agent; S3. Based on the tasks assigned by the collaboration center and fragments in the business opportunity knowledge graph, the intention understanding sub-agent analyzes the user's psychology and generates deeply persuasive content. S4. Based on the generated content, a multi-modal evaluation agent performs multi-dimensional quantitative scoring and feeds the realism evaluation results back to the creation end in real time. S5. Based on the optimized content and preset strategies, the system integrates a real-time risk control intelligent agent to perform release execution and full-process risk monitoring and intervention. S6. Based on the marketing performance data collected across the entire value chain, optimize and iterate through an evolutionary learning agent.

2. The cross-platform content marketing method based on multi-agent collaboration as described in claim 1, characterized in that: In step S1, the system continuously collects fragmented data from the platform via a crawler interface and the platform's open API. The collected raw data is cleaned, and key entity identification and preliminary relationship labeling are performed using natural language processing. Unstructured text is transformed into semi-structured data. Business feature vectors are extracted from the pre-processed data using a pre-trained model and custom rules. These feature vectors are then input into an association rule mining and pre-defined graph neural network model to calculate the implicit association strength between different feature entities. The mined strong association patterns are organized into entity-relationship-attribute triples and stored in a graph database to form a dynamic business opportunity knowledge graph. Entities represent business elements, relationships represent the business logic connections between them, and attributes record popularity, confidence, and timestamps. The system establishes real-time data stream monitoring, processes new data streams, and updates the dynamic business opportunity knowledge graph incrementally. Based on user interaction data, the weights of existing nodes and relationships in the dynamic business opportunity knowledge graph are dynamically adjusted.

3. The cross-platform content marketing method based on multi-agent collaboration according to claim 1, characterized in that: In S2, the collaborative central agent decomposes marketing objectives into atomic-level subtasks based on the real-time attributes of nodes in the dynamic business opportunity knowledge graph; the subtasks are described in a structured task description and published to the dynamic task market inside the system, and the task market maintains the task queue and agent state pool in real time.

4. The cross-platform content marketing method based on multi-agent collaboration according to claim 3, characterized in that: In S2, each functional intelligent agent, based on its own predefined AgentSkill library, dynamically bids for sub-tasks in the task market, and the benefit evaluation of the collaboration center performs quantitative scoring, thus achieving the following: ， In the formula, E represents the total benefit score, and M represents the dynamic skill matching degree. This represents the skill matching weighting coefficient, and L represents the load balancing degree. This represents the load balancing weighting coefficient; All bids are ranked according to the evaluation results. The scheme with the highest total benefit score is selected, and a task contract is signed with the bidding agent. During the task execution, the collaboration center continuously monitors the process. After the task is completed, the central collaborative agent verifies the output of each agent and integrates it into a structured collaborative output package.

5. The cross-platform content marketing method based on multi-agent collaboration according to claim 1, characterized in that: In step S3, the structured tasks and related dynamic business opportunity knowledge graph fragments allocated by the collaboration center are received. Based on a pre-trained intent recognition model, the fragments are deeply analyzed, and the intent confidence is calculated. This is achieved as follows: ， In the formula, Indicates confidence level of intent. The value of 'e' represents the weight of user pain points, and 'e' represents the sentiment score. It represents minute constants, identifies core user intent types, and outputs an intent parsing report.

6. The cross-platform content marketing method based on multi-agent collaboration according to claim 5, characterized in that: In S3, the intent understanding sub-agent constructs structured content instructions based on the intent parsing report. The internal prompt word engineering dynamically constructs optimal and highly specific prompt words according to the instructions, driving the model to create content. The generated initial draft of content undergoes a quick self-check within the creation agent. If it does not meet the standards, feedback is provided for parameter adjustment and regeneration.

7. The cross-platform content marketing method based on multi-agent collaboration according to claim 1, characterized in that: In step S4, the generated content is acquired, semantic features are extracted from the text, visual features are extracted from the image, and a text dimension realism score is calculated by analyzing the text features. This process is as follows: ， In the formula, This indicates the text's level of realism rating. Indicates colloquial weighting, Indicates the degree of colloquialism in the text. Indicates the weight of emotional fluctuations. Indicates the intensity of emotional fluctuations in the text. Represents a small constant; By identifying scene elements and the unintentionality of composition in an image, a visual dimension realism score is generated, achieving the following: ， In the formula, Indicates the visual realism score. Indicates the weight of visual naturalness. Indicates the intensity of visual naturalness. Indicates scene weight, Scene coverage This represents a small constant.

8. The cross-platform content marketing method based on multi-agent collaboration according to claim 1, characterized in that: In step S4, the text-dimensional realism score and the visual-dimensional score are dynamically weighted and integrated to output a comprehensive realism index, which is achieved as follows: ， In the formula, To convey a sense of overall realism, This indicates the consistency score between the text and images. Indicates the dynamic weight of the dimension; The overall realism index is returned to the content generation process in real time; if the index is lower than the system's preset dynamic threshold, the content optimization process is triggered; if the index meets the standard, the content directly enters the publishing process.

9. The cross-platform content marketing method based on multi-agent collaboration according to claim 1, characterized in that: In step S5, the optimized content and preset release strategy are obtained. Before the official release, the real-time risk control agent initiates a pre-release review, calls the compliance library synchronized with the platform rules, performs a final round of compliance and security verification on the content, and checks for conflicts in the release strategy. After passing the pre-release review, the system releases the content to the target platform according to the strategy. At the same time, the system embeds full-link monitoring points into the content to track core indicators after release. The real-time risk control agent starts simultaneously and begins listening to the data stream. After release, the real-time risk control agent enters a continuous monitoring state, collects the content's performance data in real time, and performs correlation analysis in conjunction with the current state of the dynamic business opportunity knowledge graph. Through a preset risk model, the real-time risk score of the content is dynamically calculated. When the risk score exceeds a specific threshold, the risk control agent automatically matches and executes the corresponding SOP from the preset intervention strategy library according to the risk level and type.

10. The cross-platform content marketing system based on multi-agent collaboration, implemented according to the cross-platform content marketing method based on multi-agent collaboration as described in claim 1, is characterized in that: It includes a dynamic business opportunity knowledge graph construction module, a multi-agent dynamic collaboration scheduling module, an intent-driven content generation module, a multimodal realism assessment module, a full-link real-time risk control execution module, and a strategy evolution optimization module.

Images