A Multi-Agent Collaborative Product Story Advertising Script Generation System and Method

Abstract

A multi-agent collaborative product storyline advertisement script generation system and method, relating to the field of computer data processing technology, includes: constructing a multi-agent collaborative platform; generating a knowledge graph of a vertical knowledge base for the multi-agents; periodically updating and supplementing the knowledge graph; and training and updating the multi-agents; acquiring simulated script aggregation centers corresponding to several different clustering features and global constraints of different brand highlights and selling points on each other agent; retraining each other agent based on the global constraints; generating product storyline advertisement scripts based on product information and the multi-agents; evaluating the quality of the product storyline advertisement scripts; and determining whether to trigger periodic updates based on the quality evaluation. This system achieves fully automated production of advertisement scripts through a five-level agent chain architecture, significantly improving efficiency. Simultaneously, through a dynamic integration mechanism of brand highlights and selling points, it improves content quality and conversion rates.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, specifically to a multi-agent collaborative product storyline advertising script generation system and method. Background Technology

[0002] Chinese Patent Publication No. CN118761383A discloses a method and system for generating literary scripts based on a large language model, including: obtaining the user's requirements for generating literary scripts; determining the character characteristics of multiple characters and the perspective plot of each character based on the requirements of the literary scripts using a pre-trained large language model; and generating literary scripts based on the character characteristics and perspective plot of each character.

[0003] Chinese patent CN119962496A discloses a method and system for AI script generation and element decomposition based on a large language model. In script generation, users input a brief overview of the story, along with key information such as story type, tone, and structure. Combined with inspiration and suggestions provided by the large language model, the system quickly builds the story's framework and gradually refines it into a complete story. Users can flexibly revise the script, achieving efficient human-machine collaboration. In script decomposition, the system utilizes the large language model and preset prompts to automatically decompose elements such as characters, scenes, props, and costumes from the script, forming detailed descriptions. Illegal input and redundant information are removed, and a standardized Excel spreadsheet file is finally output, facilitating efficient management and utilization by the production team.

[0004] Currently, there are two main technical solutions in the field of product storyboard advertising script generation, both of which face significant drawbacks. The first is the manual assembly line model, which relies on linear collaboration among professionals such as screenwriters, storyboard artists, and art designers. This model has three core pain points: First, the production cycle is lengthy, with a single script taking an average of 72 hours, requiring multiple cross-departmental communications from brand needs analysis to final script delivery; second, there is a severe disconnect between professional fields, for example, storyboard artists lacking materials science knowledge cannot understand that "cooling fiber technology" needs to be visualized through infrared thermal imaging comparison lenses, resulting in the technical selling point being simplified to changes in character facial expressions (such as "complexion improves"); third, the cost of modification is extremely high, if the brand requests to add a close-up shot of a "3-second cooling experiment" during the storyboard stage, the entire process must be reworked from the beginning of script rewriting, resulting in a waste of resources.

[0005] The second type is a discrete single-module AI interconnected solution, which attempts to optimize the process through technical means. Its architecture adopts a unidirectional data flow: the brand analysis module extracts product keywords (such as "cooling efficiency = 5℃ / 3s") and inputs them into the plot generation model (such as GPT). The plot output is then cut by a storyboard splitting tool according to timestamps, and finally rendered by the image generation API (such as DALL-E). However, this solution has the following fundamental technical defects: because the storyboard tool lacks a vertical product knowledge base, the final image generation module can only rely on its own training data preferences to output panoramic shots, resulting in poor visualization of core selling points. At the same time, there is a lack of reverse negotiation mechanism between modules. When the brand analyst detects that a competitor is focusing on the "lightweight" selling point, the existing interconnected architecture cannot optimize the script locally. For example, if a "fiber breathability experiment" segment needs to be inserted, the entire process must be restarted (starting from plot rewriting), and incremental adjustments cannot be achieved. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention aims to provide a method for generating product storyline advertisement scripts through multi-agent collaboration, comprising the following steps:

[0007] Step s1: Construct a multi-agent collaborative platform, generate a knowledge graph of the vertical knowledge base of the multi-agent, and regularly supplement and update the knowledge graph and train and update the multi-agent;

[0008] Step s2: Perform simulation tests and aggregation on the other agents in the multi-agent training and update process, except for the brand analysis agent, to obtain the simulation script aggregation centers corresponding to several different clustering features;

[0009] Step s3: Based on the simulation script aggregation centers corresponding to several different clustering features, obtain the global constraints of different brand hot spots and selling points on each other intelligent agent, and retrain each other intelligent agent based on the global constraints.

[0010] Step s4: Generate product storyline advertisement scripts based on product information and multi-agent communication, conduct quality assessments on the product storyline advertisement scripts, and determine whether to trigger scheduled updates based on the quality assessment.

[0011] Furthermore, the multi-agent collaborative platform has multiple agents, multiple vertical knowledge bases of agents, and user interaction terminals. Each vertical knowledge base is equipped with a data crawling terminal, which is used to collect professional instance data on a regular basis. The user interaction terminal is used to collect product information (product name, characteristics, target audience, and market demand, etc.) and provide feedback on product storyline advertising scripts.

[0012] Furthermore, the multi-agent system includes brand analysis agent A1, scriptwriter agent A2, storyboard analyst agent A3, photographer agent A4, and sound engineer agent A5.

[0013] Furthermore, the brand analysis agent A1 receives product information P, performs market positioning and brand analysis based on product characteristics, target audience and market trends, extracts product hot points Ph and recommends product selling points Ps;

[0014] The scriptwriter AI agent A2 receives the brand's key selling points Ph and Ps, understands and analyzes them, recommends a suitable plot theme S_theme, and writes the plot outline S_outline accordingly.

[0015] Storyboard analyst agent A3 receives the story outline S_outline, understands and breaks it down based on its professional domain knowledge, divides the story into n storyboards and generates the corresponding storyboard content sequence Bn={b1,b2,...,bn};

[0016] The photographer agent A4 receives the sequence of storyboard content Bn, processes it based on its photography expertise, and outputs the scene guidance data sequence In={i1,i2,...,in} and the video guidance data sequence Vn={v1,v2,...,vn} corresponding to each storyboard.

[0017] The sound engineer agent A5 is used to receive the storyboard content sequence Bn and generate the corresponding storyboard sound effect music sequence Mn={m1,m2,...,mn};

[0018] Finally, the story outline S_outline, storyboard content sequence Bn, visual guidance data sequence In, video guidance data sequence Vn, and sound effects and music sequence Mn are integrated to form the product story advertisement script S=(S_outline,Bn,In,Vn,Mn).

[0019] Furthermore, the process of generating a knowledge graph for a multi-agent vertical knowledge base, periodically supplementing and updating the knowledge graph, and training and updating the multi-agents includes:

[0020] Extract entities, entity attributes, and entity association rules from the vertical knowledge base of each agent. Treat entities as nodes and entity association rules as connecting edges between nodes to construct a knowledge graph of the vertical knowledge base of each agent. Treat the attributes of nodes as supplementary nodes.

[0021] When the vertical knowledge base collects new professional instance data, it extracts the entities, entity attributes, and association rules between entities from the professional instance data. The entities in the professional instance data are used as nodes, the association rules of the entities are used as connection edges between the nodes and the nodes in the knowledge graph, and the attributes of the nodes are used as supplementary nodes. The knowledge graph is then supplemented and updated based on the professional instance data.

[0022] When the knowledge graph is updated, the knowledge graph of each agent's vertical knowledge base is used as the training data for each agent, and each agent is trained and updated.

[0023] Furthermore, the process of simulating and aggregating the agents other than the brand analysis agent in the trained and updated multi-agent system to obtain the simulation script aggregation centers corresponding to several different clustering features includes:

[0024] Simulation tests were conducted on the agents other than the brand analysis agent after the training and update. Several simulation scripts were generated under different brand hot spots and selling points. The node networks corresponding to the content sequences of other agents in each simulation script were extracted in the knowledge graph (such as the node network corresponding to the plot outline S_outline in the content sequence in the knowledge graph of the scriptwriter agent, the node network corresponding to the storyboard content sequence Bn in the knowledge graph of the storyboard analyst agent, etc.).

[0025] Select a simulation script from several simulation scripts, and use the node network corresponding to another intelligent agent in the simulation script and the brand highlights and selling points corresponding to the simulation script as clustering features. Aggregate other simulation scripts in the several simulation scripts according to the aggregation features to obtain the simulation script aggregation center corresponding to the clustering features (the aggregation features of the simulation script aggregation centers are the same). Repeat the above aggregation process to obtain the simulation script aggregation centers corresponding to several different clustering features.

[0026] Furthermore, the process of obtaining the global constraints of different brand highlights and selling points on various other intelligent agents based on the simulation script aggregation centers corresponding to several different clustering features includes:

[0027] The quality of each simulated script in the simulated script aggregation center corresponding to different clustering features is evaluated to obtain the quality evaluation level of each simulated script in the simulated script aggregation center corresponding to different clustering features. Based on the quality evaluation level, the weight coefficients of different node networks corresponding to other agents under different brand highlights and selling points are obtained. Specifically, the quality evaluation levels of each simulated script in the simulated script aggregation center corresponding to different clustering features are summed and averaged to obtain the average quality evaluation level of the simulated scripts in the simulated script aggregation center corresponding to different clustering features. The average quality evaluation level is then normalized and converted into weight coefficients.

[0028] The weight coefficients of different node networks corresponding to other intelligent agents under different brand highlights and selling points are marked as global constraints of different brand highlights and selling points on other intelligent agents.

[0029] Furthermore, the process of retraining each other agent based on global constraints includes:

[0030] Based on the global constraints of different brand highlights and selling points on other intelligent agents, the weight coefficients of the connecting edges in the knowledge graph of the vertical knowledge base of other intelligent agents are labeled and updated.

[0031] The knowledge graph updated with the weight coefficients of the vertical knowledge bases of other agents is used as the training data for the agents. The other agents are then retrained. The training data is divided into a training set and a test set. The training set is input into the other agents for training until the loss function is stable. The model parameters are saved. The other agents are then tested using the test set until they meet the preset requirements. The trained agents are then obtained.

[0032] During training, the weights are continuously updated using the backpropagation algorithm, causing the loss function to gradually decrease until it reaches a stable state. During this period, techniques such as early stopping are used to avoid overfitting. In addition to the basic training process, various model parameters, including the learning rate, batch size, and regularization coefficient, are fine-tuned using grid search.

[0033] It should be further explained that, in the specific implementation process, the training objective function for other intelligent agents is:

[0034] ;

[0035] in, For intelligent agents The training data, , For input-output pairs (e.g., A3's input is "rapid freezing" and its output is "cold fog effect"); For the agent's prediction function, For model parameters, The weight coefficients of the node network to which the node path from input x to output y belongs;

[0036] Once the model training is complete and the parameters have been tuned, a final evaluation is performed using a test set to obtain the model's evaluation results. These results include classification metrics such as accuracy, recall, and F1 score. Based on the evaluation results on the test set, it is determined whether the model has met the expected standards. If the requirements are met, the model parameters are saved and deployment is prepared; otherwise, it is necessary to return to a previous stage to re-examine issues such as data quality, model structure, or training strategy.

[0037] Furthermore, the process of generating product-related narrative advertisement scripts based on product information and multi-agent interaction includes:

[0038] The product information collected from the user interaction terminal is input into the brand analysis agent to obtain the brand's highlights and selling points. The brand's highlights and selling points are then input into other agents. Based on the content sequence output by other agents (story outline S_outline, storyboard content sequence Bn, visual guidance data sequence In, video guidance data sequence Vn, and sound effects and music sequence Mn), a product storyline advertisement script S=(S_outline,Bn,In,Vn,Mn) is generated.

[0039] Furthermore, the process of conducting a quality assessment of the product storyline advertisement script and determining whether to trigger a scheduled update based on the quality assessment includes:

[0040] The product storyline advertisement script is subjected to three-dimensional evaluation index feature extraction to obtain the three-dimensional evaluation index of the product storyline advertisement script, the index weight of the three-dimensional evaluation index is set, and the membership matrix of the product storyline advertisement script to the preset quality evaluation level is obtained through fuzzy comprehensive evaluation.

[0041] The quality assessment level of the product storyline advertisement script is obtained through the membership matrix and indicator weights. A preset quality assessment level threshold is set. If the quality assessment level is greater than or equal to the quality assessment level threshold, the product storyline advertisement script is sent to the user interaction terminal. If the quality assessment level is less than the quality assessment level threshold, the data crawler is automatically triggered to collect professional instance data. When the vertical knowledge base collects new professional instance data, the knowledge graph is supplemented and updated according to the professional instance data. When the knowledge graph is completed, the knowledge graph of the vertical knowledge base of each agent is used as the training data for each agent to train and update each agent.

[0042] A multi-agent collaborative product story advertisement script generation system includes a multi-agent collaborative platform, which is communicatively connected to a knowledge graph module, a data processing module, a constraint training module, and a script generation feedback module.

[0043] The knowledge graph module is used to generate knowledge graphs for vertical knowledge bases of multi-agent systems, and to periodically supplement and update the knowledge graphs as well as to train and update the multi-agent systems.

[0044] The data processing module is used to simulate and aggregate the agents other than the brand analysis agent in the multi-agent training and update, and to obtain the simulation script aggregation center corresponding to several different clustering features.

[0045] The constraint training module is used to obtain the global constraints of different brand hot spots and selling points on each other intelligent agent based on the simulation script aggregation center corresponding to several different clustering features, and to retrain each other intelligent agent based on the global constraints.

[0046] The script generation feedback module is used to generate product storyline advertisement scripts based on product information and multiple agents, evaluate the quality of the product storyline advertisement scripts, and determine whether to trigger scheduled updates based on the quality evaluation.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] 1. Professional Knowledge Integration and Dynamic Evolution Capability: By extracting entities, attributes, and association rules from the vertical knowledge bases of various intelligent agents, a knowledge graph is constructed. Professional instance data (such as the latest advertising storyboard cases and photography technical standards) is collected regularly through data crawling to achieve continuous iteration of the knowledge graph. For example, when a new trend of "visualizing environmentally friendly materials" emerges in the market, the knowledge graph of the storyboard analyst intelligent agent can automatically supplement the rule "biodegradable materials → plant light and shadow mapping," solving the problem of insufficient vertical knowledge and poor visualization of selling points in existing AI modules (such as the simplification of "cooling fibers" in sun-protective clothing to facial expression changes in traditional solutions).

[0049] 2. Deep Integration of Professional Knowledge Through Multi-Agent Collaboration: The key selling points and highlights extracted by the brand analysis agent are transmitted through knowledge graph weights, driving the entire process from scriptwriting and storyboard design. Addressing the issue of fragmented processes, a five-level agent-connected architecture enables fully automated production of advertising scripts, significantly improving efficiency. Addressing the disconnect between commercial and creative aspects, a dynamic integration mechanism of brand key selling points and highlights deeply couples professional film and television expression with product marketing goals, simultaneously improving content quality and conversion rates. This ensures deep coupling between commercial elements and film and television professional standards, breaking through the bottleneck of professional field fragmentation in traditional manual models (e.g., storyboard artists lacking materials science knowledge leading to distorted selling points).

[0050] 3. Full-process automation and dynamic optimization: Through a five-level intelligent agent serial architecture (A1→A5), the entire script generation process is automated, reducing the time for generating a single script from the traditional 72 hours to minutes (e.g., the ice cup case took 2 minutes and 17 seconds). Simultaneously, simulation testing and cluster analysis pre-build a "hotspot-selling point-script structure" mapping template. When processing similar products (e.g., sun protection clothing from different brands), historically optimized weight coefficients can be directly called, avoiding redundant training and further improving efficiency by over 30%.

[0051] 4. Standardized interfaces support batch generation: Each agent's output uses standardized data interfaces (such as storyboard sequence Bn, screen guidance In), supporting end-to-end batch generation of highly consistent scripts. For example, a beverage brand can generate 10 different scripts with different focuses within minutes based on the same selling point, "icy refreshment," by adjusting the weight of selling points (such as "low sugar" and "fruity flavor"), thus overcoming the production capacity bottleneck of mass production in manual mode.

[0052] 5. A globally constrained selling point integration mechanism: Brand highlights and selling points serve as global constraints throughout the multi-agent processing. Dynamic adjustments to the weight coefficients of the knowledge graph ensure that each generation stage accurately responds to business needs. For example, after the "cooling fiber" selling point of sun-protective clothing is weighted and enhanced, the scene design automatically links to an "infrared thermal imaging comparison" lens, and the photographer configures "cool-toned close-up" parameters, increasing the selling point visualization conversion rate by 50% (compared to the blurring of selling points caused by panoramic lenses in traditional AI-connected solutions).

[0053] 6. Closed-loop optimization based on quality assessment: Three-dimensional assessment indicators (professional dimension, business dimension, and collaboration dimension) are combined with fuzzy comprehensive evaluation to quantify script quality in real time. When the quality assessment level is lower than the threshold (e.g., the predicted click-through rate is less than the industry average), knowledge graph updates and agent retraining are automatically triggered. For example, a cosmetics script triggered a knowledge graph update for the storyboard agent due to "insufficient visualization of moisturizing selling points," adding a mapping rule for "dynamic water droplet penetration effect," which subsequently improved the script's conversion rate by 35%.

[0054] 7. Dynamic Response to Brand-Specific Needs: Through simulated script clustering analysis, the system can automatically identify the optimal generation path corresponding to different brand highlights. For example, the "luxury" highlight of high-end brands and the "value for money" highlight of fast-moving consumer goods correspond to different storyboard rhythms (slow motion vs. fast cut) and sound effect styles (symphony vs. pop music), solving the brand tone misalignment problem caused by the "one-size-fits-all" generation mode in traditional solutions.

[0055] 8. Adaptive Evolution of Market Trends: The timely updating of the knowledge graph and the retraining mechanism of the intelligent agent enable the system to quickly absorb emerging technologies and consumer trends. For example, when "interactive storylines" became popular on short video platforms, the system automatically updated the narrative template of the script intelligent agent by crawling the latest cases and added "user choice branch" generation rules to ensure that the script always conforms to the platform's algorithm preferences and user aesthetic trends. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating a method for generating product storyline advertisement scripts through multi-agent collaboration, according to an embodiment of this application.

[0057] Figure 2This is a schematic diagram of a multi-agent collaborative product story advertising script generation system according to an embodiment of this application. Detailed Implementation

[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] like Figure 1 As shown, a method for generating product storyline advertisement scripts through multi-agent collaboration includes the following steps:

[0060] Step s1: Construct a multi-agent collaborative platform, generate a knowledge graph of the vertical knowledge base of the multi-agent, and regularly supplement and update the knowledge graph and train and update the multi-agent;

[0061] Step s2: Perform simulation tests and aggregation on the other agents in the multi-agent training and update process, except for the brand analysis agent, to obtain the simulation script aggregation centers corresponding to several different clustering features;

[0062] Step s3: Based on the simulation script aggregation centers corresponding to several different clustering features, obtain the global constraints of different brand hot spots and selling points on each other intelligent agent, and retrain each other intelligent agent based on the global constraints.

[0063] Step s4: Generate product storyline advertisement scripts based on product information and multi-agent communication, conduct quality assessments on the product storyline advertisement scripts, and determine whether to trigger scheduled updates based on the quality assessment.

[0064] It should be further explained that, in the specific implementation process, the multi-agent collaborative platform has multiple agents, multiple vertical knowledge bases of agents, and user interaction terminals for communication. Each vertical knowledge base is equipped with a data crawling terminal, which is used to collect professional instance data on a regular basis. The user interaction terminal is used to collect product information (product name, characteristics, target audience, market demand, etc.) and provide feedback on product storyline advertising scripts.

[0065] It should be further explained that, in the specific implementation process, the multi-agent system includes brand analysis agent A1, scriptwriter agent A2, storyboard analyst agent A3, cinematographer agent A4, and sound engineer agent A5.

[0066] It should be further explained that, in the specific implementation process, the brand analysis intelligent agent A1 is used to receive product information P, conduct market positioning and brand analysis based on product characteristics, target audience and market trends, extract the product's hot points Ph and recommend product selling points Ps;

[0067] The scriptwriter AI agent A2 receives the brand's key selling points Ph and Ps, understands and analyzes them, recommends a suitable plot theme S_theme, and writes the plot outline S_outline accordingly.

[0068] Storyboard analyst agent A3 receives the story outline S_outline, understands and breaks it down based on its professional domain knowledge, divides the story into n storyboards and generates the corresponding storyboard content sequence Bn={b1,b2,...,bn};

[0069] The photographer agent A4 receives the sequence of storyboard content Bn, processes it based on its photography expertise, and outputs the scene guidance data sequence In={i1,i2,...,in} and the video guidance data sequence Vn={v1,v2,...,vn} corresponding to each storyboard.

[0070] The sound engineer agent A5 is used to receive the storyboard content sequence Bn and generate the corresponding storyboard sound effect music sequence Mn={m1,m2,...,mn};

[0071] Finally, the story outline S_outline, storyboard content sequence Bn, visual guidance data sequence In, video guidance data sequence Vn, and sound effects and music sequence Mn are integrated to form the product story advertisement script S=(S_outline,Bn,In,Vn,Mn).

[0072] For example, first, input product information P (including features such as flash freezing, environmentally friendly materials, and targeting a young customer base). Brand analysis agent A1 extracts the key point Ph="3-second chilled summer" and selling points Ps=["flash freezing", "environmentally friendly and cost-effective"]. Based on this, scriptwriter agent A2 generates the theme "Office instantly becomes an ice bar" and a plot outline, describing a scene where white-collar workers use an ice cup to instantly chill drinks, attracting the attention of their colleagues. Storyboard analyst agent A3 integrates Ph / Ps to break the story down into 5 storyboards Bn, with the core storyboard b3 designing a visual key point of "the beverage bursts into a cold mist upon entering the cup + a 3-second countdown". b5 incorporates the environmentally friendly selling point through an overhead shot of the ice cup's petal shape. Cinematographer agent A4 provides guidance based on the storyboard output (such as using ice-blue font to enhance the sense of technology in close-ups of the cold mist) and camera movement (blurring the background and focusing on the product in group shots). Sound engineer agent A5 injects "ding-dong sound + freezing hissing sound effect" into b3 to enhance auditory memory. Finally, the structured script S is generated. Scene 3 transforms Ph into ice crystal numbers and a voiceover saying "3 seconds to chill!" Scene 5 pops up the text "Reusable 500 times" in a petal-shaped screen, echoing Ps.

[0073] It should be further noted that the output of each intelligent agent adopts a standardized data interface, which supports end-to-end batch generation of highly consistent scripts.

[0074] It should be further explained that, in the specific implementation process, the process of generating a knowledge graph for a multi-agent vertical knowledge base, periodically supplementing and updating the knowledge graph, and training and updating the multi-agents includes:

[0075] Due to the specialized nature of knowledge in generating product storyline advertising scripts and the complexity of the knowledge system for generating product storyline advertising scripts, it is unacceptable in terms of time and effort to manually build a multi-agent vertical knowledge base from scratch. This embodiment utilizes open semantic knowledge bases on Internet platforms to extract reusable knowledge data in the field of product storyline advertising scripts, such as Wikidata, ConceptNet, and OpenCyc. At the same time, in order to ensure the correctness and completeness of the knowledge data, a comparative analysis of open semantic knowledge bases such as Wikidata, ConceptNet, and OpenCyc is conducted in advance based on the professional experience and knowledge of experts, and OpenCyc is finally selected as the data source for extraction.

[0076] Each agent's vertical knowledge base is strictly limited to a professional subdomain. For example, the brand analysis agent only contains market positioning knowledge, and the storyboard analyst only stores visual symbol mapping rules. Furthermore, the professional instance data in each agent's vertical knowledge base adopts a unified triplet structure (entity, relation, attribute). For example, the entity is the product name, selling point, and season; the attribute is the specific value of the season; and the relation is the "season-selling point" association rule (such as "June → demand for sun-protective clothing with cooling sensation increases").

[0077] Extract entities, entity attributes, and entity association rules from the vertical knowledge base of each agent. Treat entities as nodes and entity association rules as connecting edges between nodes to construct a knowledge graph of the vertical knowledge base of each agent. Treat the attributes of nodes as supplementary nodes.

[0078] When the vertical knowledge base collects new professional instance data, it extracts the entities, entity attributes, and association rules between entities from the professional instance data. The entities in the professional instance data are used as nodes, the association rules of the entities are used as connection edges between the nodes and the nodes in the knowledge graph, and the attributes of the nodes are used as supplementary nodes. The knowledge graph is then supplemented and updated based on the professional instance data.

[0079] When the knowledge graph is updated, the knowledge graph of each agent's vertical knowledge base is used as the training data for each agent, and each agent is trained and updated.

[0080] It should be further explained that, in the specific implementation process, the process of simulating and aggregating the agents other than the brand analysis agent in the multi-agent training and update, and obtaining the simulation script aggregation centers corresponding to several different clustering features, includes:

[0081] Simulation tests were conducted on the agents other than the brand analysis agent after the training and update. Several simulation scripts were generated under different brand hot spots and selling points. The node networks corresponding to the content sequences of other agents in each simulation script were extracted in the knowledge graph (such as the node network corresponding to the plot outline S_outline in the content sequence in the knowledge graph of the scriptwriter agent, the node network corresponding to the storyboard content sequence Bn in the knowledge graph of the storyboard analyst agent, etc.).

[0082] Select a simulation script from several simulation scripts, and use the node network corresponding to another intelligent agent in the simulation script and the brand highlights and selling points corresponding to the simulation script as clustering features. Aggregate other simulation scripts in the several simulation scripts according to the aggregation features to obtain the simulation script aggregation center corresponding to the clustering features (the aggregation features of the simulation script aggregation centers are the same). Repeat the above aggregation process to obtain the simulation script aggregation centers corresponding to several different clustering features.

[0083] It should be further explained that, in the specific implementation process, the process of obtaining the global constraints of different brand highlights and selling points on various other intelligent agents based on the simulation script aggregation centers corresponding to several different clustering features includes:

[0084] The quality of each simulated script in the simulated script aggregation center corresponding to different clustering features is evaluated to obtain the quality evaluation level of each simulated script in the simulated script aggregation center corresponding to different clustering features. Based on the quality evaluation level, the weight coefficients of different node networks corresponding to other agents under different brand highlights and selling points are obtained. Specifically, the quality evaluation levels of each simulated script in the simulated script aggregation center corresponding to different clustering features are summed and averaged to obtain the average quality evaluation level of the simulated scripts in the simulated script aggregation center corresponding to different clustering features. The average quality evaluation level is then normalized and converted into weight coefficients.

[0085] The weight coefficients of different node networks corresponding to other intelligent agents under different brand highlights and selling points are marked as global constraints of different brand highlights and selling points on other intelligent agents.

[0086] It should be further explained that, in the specific implementation process, the process of retraining each other intelligent agent based on global constraints includes:

[0087] Based on the global constraints of different brand highlights and selling points on other intelligent agents, the weight coefficients of the connecting edges in the knowledge graph of the vertical knowledge base of other intelligent agents are labeled and updated.

[0088] The knowledge graph updated with the weight coefficients of the vertical knowledge bases of other agents is used as the training data for the agents. The other agents are then retrained. The training data is divided into a training set and a test set. The training set is input into the other agents for training until the loss function is stable. The model parameters are saved. The other agents are then tested using the test set until they meet the preset requirements. The trained agents are then obtained.

[0089] During training, the weights are continuously updated using the backpropagation algorithm, causing the loss function to gradually decrease until it reaches a stable state. During this period, techniques such as early stopping are used to avoid overfitting. In addition to the basic training process, various model parameters, including the learning rate, batch size, and regularization coefficient, are fine-tuned using grid search.

[0090] It should be further explained that, in the specific implementation process, the training objective function for other intelligent agents is:

[0091] ;

[0092] in, For intelligent agents The training data, , For input-output pairs (e.g., A3's input is "rapid freezing" and its output is "cold fog effect"); For the agent's prediction function, For model parameters, The weight coefficients of the node network to which the node path from input x to output y belongs;

[0093] Once the model training is complete and the parameters have been tuned, a final evaluation is performed using a test set to obtain the model's evaluation results. These results include classification metrics such as accuracy, recall, and F1 score. Based on the evaluation results on the test set, it is determined whether the model has met the expected standards. If the requirements are met, the model parameters are saved and deployment is prepared; otherwise, it is necessary to return to a previous stage to re-examine issues such as data quality, model structure, or training strategy.

[0094] It should be further explained that, in the specific implementation process, the process of generating product storyline advertisement scripts based on product information and multi-agent interaction includes:

[0095] The product information collected from the user interaction terminal is input into the brand analysis agent to obtain the brand's highlights and selling points. The brand's highlights and selling points are then input into other agents, which output content sequences (story outline S_outline, storyboard content sequence Bn, visual guidance data sequence In, video guidance data sequence Vn, and sound effects and music sequence Mn). Based on the content sequences output by each other agent, a product storyline advertisement script S=(S_outline,Bn,In,Vn,Mn) is generated.

[0096] By replacing traditional manual collaboration with a five-level intelligent agent interconnected architecture (A1→A2→A3→A4→A5), the entire process of advertising script production is automated, improving generation efficiency by over 70%. The brand analysis intelligent agent A1 extracts the brand's key points (Ph) and selling points (Ps), which not only drive the initial script creation (A2) but also serve as a global constraint, continuously inputting and deeply integrating into the processing of all subsequent intelligent agents (A3, A4, A5). This ensures a deep unity between film and television professionalism and commercial goals. Through the dynamic integration mechanism of brand key points (Ph) and selling points (Ps), the storyboard design, visual language, and sound effects arrangement accurately respond to product positioning and customer preferences, increasing advertising conversion rates by 30%-50%.

[0097] It should be further explained that, in the specific implementation process, the process of conducting a quality assessment of the product storyline advertisement script and determining whether to trigger a scheduled update based on the quality assessment includes:

[0098] The product storyline advertisement script is subjected to three-dimensional evaluation index feature extraction to obtain the three-dimensional evaluation index of the product storyline advertisement script, the index weight of the three-dimensional evaluation index is set, and the membership matrix of the product storyline advertisement script to the preset quality evaluation level is obtained through fuzzy comprehensive evaluation.

[0099] The quality assessment level of the product storyline advertisement script is obtained through the membership matrix and indicator weights. A preset quality assessment level threshold is set. If the quality assessment level is greater than or equal to the quality assessment level threshold, the product storyline advertisement script is sent to the user interaction terminal. If the quality assessment level is less than the quality assessment level threshold, the data crawler is automatically triggered to collect professional instance data. When the vertical knowledge base collects new professional instance data, the knowledge graph is supplemented and updated according to the professional instance data. When the knowledge graph is completed, the knowledge graph of the vertical knowledge base of each agent is used as the training data for each agent to train and update each agent.

[0100] It should be further explained that, in the specific implementation process, the three-dimensional evaluation indicators include professional, business, and collaborative dimensions.

[0101] The professional dimension includes:

[0102] Accuracy of viral hits: Accuracy of viral hits = Number of viral hits verified by market research / Total number of viral hits; For example, in the case of the ice cup, the viral hit "3-second chilled summer" was verified through Douyin hot keyword analysis, with an accuracy rate of 100%;

[0103] Selling point quantification: Selling point quantification = Number of verifiable quantifiable selling points / Total number of selling points; for example, if "rapid freezing" is associated with "cooling down in 3 seconds" data, the quantification rate is 100%.

[0104] Theme fit: Theme fit = Number of key scenes in the storyline where the product's selling points and highlights appear / Total number of scenes; For example, in the ice cup case, the theme "Office instantly transforms into an ice bar" appears 3 times in 5 panels, so the fit = 60%;

[0105] Storyboard grammar compliance rate: Based on film and television industry standards, the storyboard duration allocation and shot switching logic are checked. Storyboard grammar compliance rate = number of shots that conform to storyboard grammar / total number of shots.

[0106] Photographic parameter matching degree: Compare with the parameter database of professional photography textbooks to calculate the lens parameter matching degree. Photographic parameter matching degree = ∑(cosine similarity between actual parameters and standard parameters) / total number of parameters;

[0107] Audio-visual synchronization accuracy: Synchronization error = ∑ | Sound effect trigger time - Screen event time | / Total number of sound effects;

[0108] The business dimension includes:

[0109] Embedding Awkwardness: The abruptness of mentioning selling points in the dialogue is detected using an NLP sentiment analysis model. Embedding Awkwardness = 1 - Selling Point Context Semantic Coherence Score / Full Score;

[0110] Audience profile matching degree: Calculate the overlap of plot elements based on target audience tags (such as age, spending power). Audience profile matching degree = number of plot elements that match the audience's preferences / total number of elements;

[0111] Collaboration dimensions include:

[0112] Completeness of Product Highlights and Selling Points (Persistence Rate): This is checked by traversing the knowledge graph. Persistence rate = number of product highlights and selling points used by other agents / total number of product highlights and selling points output by brand analysis agent A1. For example, in the ice cup case, both product selling points are passed through A3-A5, resulting in a persistence rate of 100%.

[0113] Storyboard Mapping Degree: Storyboard Mapping Degree = Number of Storyboard Scenes Recreated in the Storyboard / Total Number of Story Scenes; Example: Ice Cup has 5 storyboard scenes that completely recreate the story outline, so the mapping degree is 100%.

[0114] It should be further explained that, in the specific implementation process, the process of obtaining the quality assessment level of the product storyline advertisement script based on the membership matrix and indicator weights includes:

[0115] The evaluation index weights and membership matrix of the evaluation index are fused by formula to obtain the fuzzy comprehensive evaluation matrix of the evaluation index. The membership degree of the product plot advertisement script to different quality evaluation levels is obtained according to the fuzzy comprehensive evaluation matrix. The quality evaluation level with the highest membership degree corresponding to the product plot advertisement script is selected and the quality evaluation level with the highest membership degree corresponding to the product plot advertisement script is taken as the quality evaluation level of the product plot advertisement script.

[0116] The formula is as follows:

[0117] ;

[0118] in, The fuzzy comprehensive evaluation matrix of the evaluation indicators is... The weights of the evaluation indicators are: For the membership matrix, " "" indicates that the elements at corresponding positions in the weight matrix and membership matrix of the evaluation index are multiplied together. The weighting parameter is used to control the balance between the weight matrix and the membership matrix in the fuzzy comprehensive evaluation matrix of the evaluation index.

[0119] like Figure 2As shown, a multi-agent collaborative product story advertisement script generation system includes a multi-agent collaborative platform, which is communicatively connected to a knowledge graph module, a data processing module, a constraint training module, and a script generation feedback module.

[0120] The knowledge graph module is used to generate knowledge graphs for vertical knowledge bases of multi-agent systems, and to periodically supplement and update the knowledge graphs as well as to train and update the multi-agent systems.

[0121] The data processing module is used to simulate and aggregate the agents other than the brand analysis agent in the multi-agent training and update, and to obtain the simulation script aggregation center corresponding to several different clustering features.

[0122] The constraint training module is used to obtain the global constraints of different brand hot spots and selling points on each other intelligent agent based on the simulation script aggregation center corresponding to several different clustering features, and to retrain each other intelligent agent based on the global constraints.

[0123] The script generation feedback module is used to generate product storyline advertisement scripts based on product information and multiple agents, evaluate the quality of the product storyline advertisement scripts, and determine whether to trigger scheduled updates based on the quality evaluation.

[0124] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for generating a commodity scenario advertisement script in multi-agent cooperation, characterized in that, The method comprises the following steps: Step s1: constructing a multi-agent collaborative platform, generating a knowledge graph of the vertical knowledge base of the multi-agent, performing periodic supplementary updating on the knowledge graph, and training and updating the multi-agent; Step s2: performing simulation testing on the other agents except the brand analysis agent after training and updating, generating a plurality of simulation scripts corresponding to different brand highlights and selling points, and extracting the node network corresponding to the content sequence of the other agents in each simulation script in the knowledge graph; Selecting a simulation script from the plurality of simulation scripts, taking the node network corresponding to the other agent in the simulation script and the brand highlight and selling point corresponding to the simulation script as clustering features, aggregating the other simulation scripts in the plurality of simulation scripts according to the clustering features, obtaining the simulation script aggregation center corresponding to the clustering features, and repeating the above aggregation process to obtain the simulation script aggregation centers corresponding to a plurality of different clustering features; Step s3: obtaining the quality evaluation level of each simulation script in the simulation script aggregation center corresponding to different clustering features, and obtaining the weight coefficient of the different node networks corresponding to each other agent under different brand highlight and selling point conditions based on the quality evaluation level; Marking the weight coefficients of the different node networks corresponding to each other agent under different brand highlight and selling point conditions as global constraint conditions of different brand highlights and selling points on each other agent, and retraining each other agent based on the global constraint conditions; Step s4: generating a product plot advertisement script according to the product information and the multi-agent, performing quality evaluation on the product plot advertisement script, and determining whether to trigger periodic supplementary updating according to the quality evaluation. 2.The method of claim 1, wherein, The multi-agent collaborative platform is communicatively connected with a plurality of agents, a plurality of vertical knowledge bases of the agents, and a user interaction terminal, and each vertical knowledge base is provided with a data crawling terminal for periodically collecting professional example data, and the user interaction terminal is used for collecting product information and feeding back product plot advertisement scripts. 3.The method of claim 2, wherein, The multi-agent includes a brand analysis agent, a scriptwriter agent, a split-screen analyst agent, a photographer agent, and a sound engineer agent.

4. The method of claim 3, wherein the method further comprises: The process of generating the knowledge graph of the vertical knowledge base of the multi-agent, periodically supplementing the knowledge graph, and training and updating the multi-agent comprises: Extracting entities, attributes of the entities, and association rules of the entities in the vertical knowledge base of each agent, taking the entities as nodes, taking the association rules of the entities as connection edges between the nodes, constructing the knowledge graph of the vertical knowledge base of each agent, and taking the attributes of the nodes as supplementary nodes of the nodes; When the vertical knowledge base collects new professional example data, the knowledge graph is supplemented and updated according to the professional example data; When the knowledge graph is completed, the knowledge graph of the vertical knowledge base of each agent is taken as the training data of each agent, and each agent is trained and updated.

5. The multi-agent collaborative commercial scenario advertisement script generation method according to claim 4, characterized in that, The connection edges in the knowledge graph of each other agent are updated with weight coefficients according to the global constraint conditions of different brand highlights and selling points on each other agent; The other agents are retrained by taking the updated knowledge graph with the weight coefficients of the other agents as training data of the agents, to obtain the trained other agents.

6. The multi-agent collaborative commercial scenario advertisement script generation method according to claim 5, characterized in that, The process of generating the product plot advertisement script includes: The brand analysis agent is input with the product information, brand highlights and selling points are obtained, the brand highlights and selling points are input into the other agents, the other agents output content sequences, and the product plot advertisement script is generated based on the content sequences output by the other agents.

7. The method of claim 6, wherein the method further comprises: The process of quality evaluation of the product plot advertisement script and determination of whether to trigger the timed supplementary update according to the quality evaluation includes: Three-dimensional evaluation index feature extraction is performed on the product plot advertisement script to obtain three-dimensional evaluation indexes of the product plot advertisement script, index weights of the three-dimensional evaluation indexes are set, and a membership matrix of the product plot advertisement script for a preset quality evaluation level is obtained through fuzzy comprehensive evaluation; The quality evaluation level of the product plot advertisement script is obtained through the membership matrix and the index weights, a preset quality evaluation level threshold is set, if the quality evaluation level is greater than or equal to the quality evaluation level threshold, the product plot advertisement script is sent to the user interaction end, and if the quality evaluation level is less than the quality evaluation level threshold, professional example data is automatically collected by the data crawling end.

8. A multi-agent collaborative commercial plot advertisement script generation system, specifically applied to the multi-agent collaborative commercial plot advertisement script generation method of any one of claims 1 to 7, characterized in that, The multi-agent collaborative platform is connected in communication with a knowledge graph module, a data processing module, a constraint training module and a script generation feedback module; The knowledge graph module is used to generate a knowledge graph of a vertical knowledge base of the multi-agent, to perform timed supplementary update on the knowledge graph and to perform training update on the multi-agent; The data processing module is used to perform simulation test on the other agents except the brand analysis agent after training update, to generate a plurality of simulation scripts corresponding to different brand highlights and selling points, and to extract node networks corresponding to the content sequences of the other agents in the simulation scripts in the knowledge graph; A certain simulation script is selected from the plurality of simulation scripts, the node networks corresponding to a certain other agent in the certain simulation script and the brand highlights and selling points corresponding to the certain simulation script are taken as clustering features, other simulation scripts in the plurality of simulation scripts are aggregated according to the clustering features, a simulation script aggregation center corresponding to the clustering features is obtained, the above aggregation process is repeated, and simulation script aggregation centers corresponding to a plurality of different clustering features are obtained; The constraint training module is used to obtain weight coefficients of different node networks corresponding to the other agents under different brand highlights and selling points based on quality evaluation levels of the simulation scripts in the simulation script aggregation centers corresponding to the different clustering features; The weight coefficients of the different node networks corresponding to the other agents under different brand highlights and selling points are marked as global constraint conditions of the different brand highlights and selling points on the other agents, and the other agents are retrained based on the global constraint conditions; The script generation feedback module is used to generate a product plot advertisement script according to product information and the multi-agent, to perform quality evaluation on the product plot advertisement script, and to determine whether to trigger the timed supplementary update according to the quality evaluation.

Images