Feedback attribution-based short play advertisement generation model optimization method and system

Abstract

This application relates to the field of advertising delivery technology, and in particular to a method and system for optimizing a short drama advertising generation model based on feedback attribution. The method includes: acquiring task-related information; constructing a demand state vector and generating a hierarchical narrative state diagram; determining corresponding candidate narrative paths; calling the short drama advertising generation model to generate candidate advertising scripts and forming a candidate advertising version set; conducting delivery tests on the candidate advertising versions in the candidate advertising version set and collecting raw feedback data corresponding to each candidate advertising version; performing feedback attribution and denoising processing on the raw feedback data to obtain attribution feedback results corresponding to each candidate advertising version; constructing preference samples among the candidate advertising versions based on the attribution feedback results; and optimizing and updating the short drama advertising generation model based on the preference samples and the narrative structure constraints corresponding to the hierarchical narrative state diagram. This application can continuously and stably output short drama advertising content with a complete structure and better meeting the requirements of delivery effectiveness.

Description

Technical Field

[0001] This application relates to the field of advertising technology, and in particular to an optimization method and system for a short drama advertising generation model based on feedback attribution. Background Technology

[0002] In existing technologies, short drama ad script generation is typically accomplished by a script generation model. Specifically, after receiving information such as advertising requirements, product selling points, target audience, brand rules, and platform requirements, the system uses this information as input to the script generation model. The model then generates multiple candidate short drama ad scripts, which typically include opening content, plot progression content, product or service descriptions, and action guidance content. Once generated, the system can create corresponding ad versions based on the candidate scripts and conduct small-scale testing on different ad versions, collecting feedback metrics such as impressions, click-through rates, conversion rates, viewing duration, and sales data. For ad versions with good performance, existing solutions typically use their corresponding script content as superior samples and the script content corresponding to ad versions with poor performance as inferior samples. These samples are then used to train or adjust the parameters of the script generation model, causing it to tend to output script content similar to the high-performing ad versions in subsequent generation processes.

[0003] However, existing script generation models typically generate candidate short drama ad scripts directly based on advertising needs. While this can create multiple script versions, the generation process relies heavily on the model's overall understanding of the input text, lacking clear constraints on the internal content units and their sequential relationships within the short drama ad. This can lead to instability in the transitions between opening attraction, plot progression, product or service descriptions, credible information presentation, and action guidance, resulting in fluctuations in narrative completeness and content coherence among candidate scripts. Furthermore, when updating the script generation model using small-scale campaign results, existing solutions often rely directly on raw feedback metrics such as impressions, click-through rates, conversion rates, viewing duration, and transaction data to judge the quality of ad versions. These metrics are also influenced by factors such as campaign time, placement, budget, audience segmentation, platform traffic allocation, and creative presentation. Directly using the overall feedback of the ad version as the basis for model updates can easily lead to misinterpreting occasional differences in external campaign conditions as differences in the quality of the script content itself. Therefore, existing short drama ad generation models struggle to stably control the narrative structure of the script content and accurately update based on reliable feedback, thus affecting the consistent improvement of subsequent short drama ad generation quality. Therefore, how to balance script narrative structure control and reliable utilization of delivery feedback during the optimization of short drama ad generation models, so that the generation models can continuously and stably output short drama ad content with complete structure and better meet the needs of delivery effect, has become a problem that needs to be solved. Summary of the Invention

[0004] This application provides a method and system for optimizing a short drama ad generation model based on feedback attribution, which can continuously and stably output short drama ad content with a complete structure and better meeting the needs of ad delivery performance. This application provides the following technical solutions: Firstly, this application provides an optimization method for a short drama advertisement generation model based on feedback attribution, the method comprising: Obtain task association information for short drama ad generation tasks, construct a demand state vector based on the task association information, and generate a hierarchical narrative state diagram based on the demand state vector to describe the ad narrative unit and its transition relationship, and determine the candidate narrative path corresponding to the hierarchical narrative state diagram. A preset short drama ad generation model is invoked to generate candidate ad scripts based on the demand state vector, the hierarchical narrative state diagram, and the candidate narrative path, and a candidate ad version set is formed based on the candidate ad scripts; Test the candidate ad versions in the candidate ad version set and collect the original feedback data corresponding to each candidate ad version; perform feedback attribution and noise reduction processing on the original feedback data to obtain the attribution feedback results corresponding to each candidate ad version. Based on the attribution feedback results, a preference sample is constructed among the candidate ad versions, and the preset short drama ad generation model is optimized and updated based on the preference sample and the narrative structure constraints corresponding to the hierarchical narrative state diagram.

[0005] In one specific implementation, the step of generating a hierarchical narrative state diagram based on the demand state vector to describe advertising narrative units and their transition relationships, and determining candidate narrative paths corresponding to the hierarchical narrative state diagram, includes: A hierarchical narrative state diagram is generated based on a preset narrative unit library, which includes narrative character nodes and fragment instance nodes, wherein the fragment instance nodes are specific content fragments under each narrative character node. Based on the product category, target audience, brand rules, and platform constraints determined in the task association information, and in conjunction with the demand state vector, the fragment instance nodes in the preset narrative unit library are filtered, the filtered fragment instance nodes are determined as candidate narrative nodes, and the candidate narrative nodes are connected to their respective narrative role nodes. When the relationship between the narrative roles of two candidate narrative nodes conforms to the preset narrative order rules, and there are no conflicts in content tags, brand rules, or platform constraints between them, a directed transfer edge is established between the two candidate narrative nodes. From the hierarchical narrative state diagram, a candidate narrative path is determined with the candidate narrative node corresponding to the opening attraction node as the path start point, the candidate narrative node corresponding to the action guidance node as the path end point, and the candidate narrative nodes corresponding to the user pain point node, product or service description node and credible information node are sequentially included between the path start point and the path end point.

[0006] In a specific feasible implementation, the step of invoking a preset short drama advertisement generation model, generating candidate advertisement scripts based on the demand state vector, the hierarchical narrative state diagram, and the candidate narrative path, and forming a candidate advertisement version set based on the candidate advertisement scripts includes: A pre-trained conditional text generation model is invoked as the preset short drama advertisement generation model. The conditional text generation model includes a conditional input layer, a graph structure encoding layer, a segment state control layer, and a text generation layer. The conditional input layer receives the demand state vector, and the graph structure encoding layer encodes the narrative character nodes, candidate narrative nodes, and directed transition edges in the hierarchical narrative state graph to obtain a global graph representation. The fragment state control layer generates the fragment state vector corresponding to each script content segment according to the arrangement order of the candidate narrative nodes in the candidate narrative path. Under the joint constraints of the demand state vector, the global graph representation, and the fragment state vector, the text generation layer generates each script content segment sequentially according to the arrangement order of the candidate narrative nodes in the candidate narrative path, and combines each script content segment to obtain the candidate advertisement script. The candidate ad scripts are verified for content unit integrity, brand rules, platform rules, and script duration. Candidate ad scripts that pass the verification are retained. Based on the retained candidate ad scripts, camera instructions, narration text, and subtitle schemes are generated. The script text, camera instructions, narration text, and subtitle schemes corresponding to the same candidate ad script are associated as candidate ad versions. Multiple candidate ad versions are aggregated to obtain a candidate ad version set.

[0007] In one specific implementation scheme, the step of testing the delivery of candidate ad versions in the candidate ad version set and collecting raw feedback data corresponding to each candidate ad version includes: Assign a version identifier to each candidate ad version in the candidate ad version set, and generate feedback records based on the exposure logs, click logs, conversion logs, viewing logs, and transaction logs recorded by the ad delivery platform; Each feedback record corresponds to one exposure sample, and each feedback record includes the candidate ad version identifier, whether it was clicked, whether it was converted, the viewing duration, the transaction amount, and the corresponding placement context. The placement context includes the placement time period, the exposure position, the budget level, the audience segment, and the placement platform. The feedback records were aggregated according to the candidate ad version identifier to obtain the original feedback data corresponding to each candidate ad version; The original feedback data is subjected to invalid record removal and abnormal record removal. The feedback records retained after noise reduction are used as valid feedback samples. The invalid record removal includes deleting feedback records that lack candidate ad version identifiers, lack delivery context, lack exposure results, or fail to transmit transaction amount. The abnormal record removal includes deleting feedback records that repeatedly trigger clicks on the same candidate ad version more than a preset number of times within a preset time window using the same anonymous session identifier or de-identified trigger identifier, as well as feedback records with zero viewing time, which cannot match the effective exposure link but have conversion results.

[0008] In one specific implementation, the step of performing feedback attribution and denoising processing on the original feedback data to obtain the attribution feedback results corresponding to each candidate ad version includes: Based on the standardized values ​​of click markers, conversion markers, view completion rate, and transaction amount corresponding to the valid feedback samples, calculate the feedback revenue observation value of the valid feedback samples; The valid feedback samples were grouped into contexts based on the time period, exposure location, budget level, audience segmentation, and advertising platform. The average of the feedback revenue observations of all valid feedback samples within the same context group was determined as the baseline revenue for the corresponding advertising context. The probability of a candidate ad version being assigned to a specific delivery context is determined by the ratio of the number of times a candidate ad version is exposed within the same context group to the total number of times all candidate ad versions within that context group are exposed. Based on the observed feedback revenue, the baseline revenue, and the propensity probability, a causal correction feedback value for the candidate ad version is calculated, and the causal correction feedback value is expressed as follows: ; in, Indicates candidate ad version The causal correction feedback value; Indicates candidate ad version The number of valid feedback samples that participate in the calculation of attribution feedback results after denoising processing, and Greater than 0; Indicates candidate ad version The corresponding number One valid feedback sample; Indicates candidate ad version The Feedback benefit observations for each valid feedback sample; Indicates the first The delivery context corresponding to each valid feedback sample; Indicates the context of the delivery. Corresponding baseline returns; Indicates the candidate ad version to be displayed. The delivery action; Indicates the context of the delivery. Next candidate ad version The probability of being assigned a specific resource; This represents the preset stability constant; The candidate ad version identifier, the number of valid feedback samples, the denoised click statistics, the denoised conversion statistics, the denoised view statistics, the denoised transaction statistics, and the causal correction feedback value are combined to obtain the attribution feedback result corresponding to the candidate ad version.

[0009] In one specific implementation, constructing a preference sample among the candidate ad versions based on the attribution feedback results includes: For any two candidate ad versions, read the causal correction feedback value and the number of valid feedback samples corresponding to the two candidate ad versions respectively; When the number of valid feedback samples for both candidate ad versions is not less than the preset minimum number of samples, and the difference between the causal correction feedback value of one candidate ad version and the causal correction feedback value of the other candidate ad version is greater than the preset feedback difference threshold, the candidate ad version with the higher causal correction feedback value is selected as the preferred version, and the candidate ad version with the lower causal correction feedback value is selected as the inferior version. The candidate ad scripts corresponding to the preferred version and the inferior version are respectively used to form a preference sample. If the number of valid feedback samples for two candidate ad versions does not reach the preset minimum number of samples, or if the difference between the causal correction feedback values ​​of the two candidate ad versions does not exceed the preset feedback difference threshold, then the two candidate ad versions will not be used to construct preference samples. A preference sample set is formed based on the constructed preference samples.

[0010] In one specific implementation scheme, optimizing and updating the preset short drama advertisement generation model based on the preference samples and the narrative structure constraints corresponding to the hierarchical narrative state diagram includes: The candidate ad scripts in the preference sample are mapped to the candidate narrative nodes in the corresponding candidate narrative path, and the structural quality score of the candidate ad scripts is determined based on the average of the node coverage ratio, the order consistency ratio, and the rule satisfaction ratio. The structural dominance difference is determined by the difference between the structural quality scores of the preferred scripts and the undesired scripts in the preference sample. ; in, Indicates the preferred script Compared to the inferior selection script Poor structural advantages; Indicates the preferred script The structural mass fraction; Indicates the script for the inferior selection The structural mass fraction; Based on the log difference in generation probabilities of the current model to be optimized for the preferred and undesired scripts, the log difference in generation probabilities of the reference model for the preferred and undesired scripts, and the aforementioned structural advantage difference, a preference optimization loss is constructed: ; in, Indicate the loss of preference optimization; Represents a set of preference samples; Represents the set of preference samples A preference sample in which candidate ad versions Corresponding preferred script Candidate Ad Versions Corresponding inferior selection script ; Represents a logarithmic function; Represents the Sigmoid function; This indicates the current model to be optimized relative to the preferred script. and inferior selection script The logarithmic difference of the generation probability; Indicates the reference model for the preferred script and inferior selection script The logarithmic difference of the generation probability; Indicates the preferred script Compared to the inferior selection script Poor structural advantages; Represents the structural constraint weights. Preset non-negative weights; With the goal of minimizing the preference optimization loss, the trainable parameters in the preset short drama advertisement generation model are updated.

[0011] Secondly, this application provides a short drama advertisement generation model optimization system based on feedback attribution, which adopts the following technical solution: A feedback attribution-based optimization system for short drama ad generation models includes: The narrative path construction module is used to obtain the task association information of the short drama advertisement generation task, construct the demand state vector according to the task association information, generate a hierarchical narrative state diagram to describe the advertisement narrative unit and its transition relationship based on the demand state vector, and determine the candidate narrative path corresponding to the hierarchical narrative state diagram. The ad version generation module is used to call a preset short drama ad generation model, generate candidate ad scripts based on the demand state vector, the hierarchical narrative state diagram and the candidate narrative path, and form a candidate ad version set based on the candidate ad scripts; The feedback attribution processing module is used to conduct delivery tests on candidate ad versions in the candidate ad version set, collect raw feedback data corresponding to each candidate ad version, and perform feedback attribution and noise reduction processing on the raw feedback data to obtain the attribution feedback results corresponding to each candidate ad version. The model optimization and update module is used to construct preference samples among the candidate ad versions based on the attribution feedback results, and to optimize and update the preset short drama ad generation model based on the preference samples and the narrative structure constraints corresponding to the hierarchical narrative state diagram.

[0012] Thirdly, this application provides an electronic device, the device including a processor and a memory; the memory stores a program, the program being loaded and executed by the processor to implement a feedback attribution-based short drama advertising generation model optimization method as described in the first aspect.

[0013] Fourthly, this application provides a computer-readable storage medium storing a program that, when executed by a processor, is used to implement a feedback attribution-based short drama advertising generation model optimization method as described in the first aspect.

[0014] By acquiring task association information of the short drama ad generation task, a demand state vector is constructed. Based on the demand state vector, a hierarchical narrative state diagram is generated to describe the ad narrative units and their transition relationships, thereby determining candidate narrative paths. This ensures that the short drama ad script forms structured constraints on the ad narrative units and their sequential connections before generation. On this basis, a pre-set short drama ad generation model is invoked to generate candidate ad scripts based on the demand state vector, hierarchical narrative state diagram, and candidate narrative paths, forming a set of candidate ad versions. This allows the candidate ad content to be organized according to a clear narrative path. Subsequently, the candidate ad versions are tested and raw feedback data is collected. Attribution feedback results are obtained through feedback attribution and noise reduction processing, so that the feedback is no longer directly based on the raw feedback data as the basis for model updates, but rather reflects the feedback performance of the candidate ad versions after attribution and noise reduction. Finally, preference samples among the candidate ad versions are constructed based on the attribution feedback results, and the pre-set short drama ad generation model is optimized and updated in conjunction with the narrative structure constraints corresponding to the hierarchical narrative state diagram. Therefore, this application, on the one hand, uses hierarchical narrative state diagrams and candidate narrative paths to structurally control the content units and their connections in the short drama advertisement script; on the other hand, it improves the reliability of the delivery feedback when participating in model optimization through attribution feedback results and preference samples. Thus, it can take into account both script narrative structure control and reliable utilization of delivery feedback in the process of optimizing the short drama advertisement generation model, so that the updated generation model can more stably output short drama advertisement content that is structurally complete, content-coherent, and more in line with the delivery effect requirements.

[0015] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the optimization method for the short drama advertising generation model based on feedback attribution in the embodiments of this application.

[0017] Figure 2 This is a structural block diagram of the short drama advertisement generation model optimization system based on feedback attribution in the embodiments of this application.

[0018] Figure 3 This is a block diagram of an electronic device optimized by a short drama advertisement generation model based on feedback attribution in an embodiment of this application. Detailed Implementation

[0019] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0020] Optionally, this application uses the feedback attribution-based short drama advertising generation model optimization method provided in various embodiments as an example for application in an electronic device. The electronic device is a terminal or a server. The terminal can be a computer, tablet computer, etc. This embodiment does not limit the type of electronic device.

[0021] Reference Figure 1 This is a flowchart illustrating a method for optimizing a short drama advertising generation model based on feedback attribution, provided in one embodiment of this application. The method includes at least the following steps: Step S101: Obtain the task association information of the short drama advertisement generation task, construct the demand state vector based on the task association information, and generate a hierarchical narrative state diagram to describe the advertisement narrative unit and its transition relationship based on the demand state vector, and determine the candidate narrative path corresponding to the hierarchical narrative state diagram.

[0022] In step S101, the task-related information includes advertising requirement text, brand rules, platform constraints, and historical audience feedback. The advertising requirement text includes basic information about the product or service to be promoted, its core selling points, target audience, and advertising objectives. Brand rules include brand tone, prohibited expressions, disclaimer rules, and compliance requirements. Platform constraints include ad duration, aspect ratio, subtitle area, and review rules. Historical audience feedback includes impressions, click-through rates, conversion rates, viewing time, and sales data for similar ads in past campaigns. When it is necessary to combine creative materials to form an ad version, the task-related information also includes an existing creative material index, which records the type, tags, and storage location of available creative materials.

[0023] It should be noted that the task-related information comes from the task submission data of the advertiser corresponding to the short drama ad generation task, the brand rule data pre-configured by the brand, the delivery specification data opened by the delivery platform to the advertiser's account, the historical delivery statistics of existing ad versions under the advertiser's account, and the material data in the advertiser's authorized material library. Specifically, the ad requirement text is submitted by the advertiser when creating the short drama ad generation task; the brand rules are pre-entered and maintained by the brand or advertiser; the delivery platform constraints are determined by the ad creation specifications, review specifications, and format specifications provided by the delivery platform to the advertiser's account; historical audience feedback comes from the aggregated statistical results of similar ad versions that have been delivered under the advertiser's account; and the existing material index comes from the material records in the advertiser's authorized material library that have been granted usage rights. Historical audience feedback is summarized by ad version, delivery batch, or audience group, and does not include personal information that can directly identify a natural person; the existing material index only records the material type, material tags, and material storage location, without changing the material ownership or authorization scope.

[0024] After obtaining the task-related information, encoding processing is performed according to the information type. The advertising demand text is semantically encoded to obtain an advertising demand text vector; brand rules are broken down into rule entries, and the rule type, rule object, and rule strength of each rule entry are encoded to obtain a brand rule vector; the advertising duration, aspect ratio, subtitle area, and review rules in the platform constraints are encoded to obtain a platform constraint vector; and the exposure, click-through rate, conversion rate, viewing time, and transaction data in historical audience feedback are statistically encoded to obtain a historical feedback vector. The advertising demand text vector, brand rule vector, platform constraint vector, and historical feedback vector are mapped to the same dimension and then concatenated in a fixed order to obtain a demand status vector. Through this processing, different types of task-related information are vectorized according to their corresponding data types, avoiding the mixing of text, rules, platform constraints, and feedback statistics.

[0025] After obtaining the demand state vector, a hierarchical narrative state diagram is generated based on a pre-set narrative unit library. The pre-set narrative unit library includes narrative role nodes and fragment instance nodes. Narrative role nodes include opening attraction nodes, user pain point nodes, product or service description nodes, credible information nodes, and action guidance nodes. Fragment instance nodes are specific content fragments under each narrative role node, and each fragment instance node is pre-configured with content tags, applicable categories, applicable audiences, recommendation duration, evidence strength, emotional intensity, compliance flags, and whether it must appear. Based on the product category, target audience, brand rules, and platform constraints determined in the task association information, and combined with the demand state vector, fragment instance nodes in the pre-set narrative unit library are screened. Fragment instance nodes that meet the requirements of applicable category matching, applicable audience matching, recommendation duration not exceeding the platform's allowed duration, and compliance flags not violating brand rules and review rules are identified as candidate narrative nodes. These candidate narrative nodes are then connected to their respective narrative role nodes, forming the node hierarchy structure of the hierarchical narrative state diagram.

[0026] After identifying candidate narrative nodes, directed transition edges are generated between them based on preset narrative order rules, brand rules, and platform constraints. Specifically, when the sequential relationship between the narrative roles of two candidate narrative nodes conforms to the preset narrative order rules, and there are no conflicts in content tags, brand rules, or platform constraints between them, a directed transition edge is established between the two candidate narrative nodes, and the starting node, target node, and connection type of the directed transition edge are recorded. The preset narrative order rules include at least the following: opening attraction at the beginning of the path, user pain points before product or service descriptions, a transition relationship between credible information and product or service descriptions, and action guidance at the end or near the end of the path. Connection types include progressive, causal, reversal, and proof relationships, determined based on the narrative roles of the starting and target nodes and their content tags. Thus, the hierarchical narrative state diagram is composed of narrative role nodes, candidate narrative nodes, and directed transition edges between candidate narrative nodes.

[0027] After generating the hierarchical narrative state diagram, candidate narrative paths are determined from it. Specifically, the candidate narrative node corresponding to the opening attraction node is used as the starting point of the path, and the candidate narrative node corresponding to the action guidance node is used as the ending point. Candidate narrative nodes corresponding to user pain point nodes, product or service description nodes, and credible information nodes are then added sequentially between the starting and ending points to form an initial narrative path. For any initial narrative path, if there are directed transition edges between adjacent candidate narrative nodes and none of the candidate narrative nodes violate brand rules or platform review rules, then the initial narrative path is determined as an optional narrative path. If there are candidate narrative nodes in the initial narrative path where no directed transition edges are established between adjacent candidate narrative nodes, or if there are candidate narrative nodes that violate brand rules or platform review rules, then the initial narrative path is eliminated.

[0028] After obtaining the optional narrative paths, path validation is performed. Path validation includes role integrity validation, duration validation, and duplicate content validation. Role integrity validation determines whether the optional narrative path contains candidate narrative nodes corresponding to opening attraction, user pain points, product or service descriptions, credible information, and action guidance; if any of these candidate narrative nodes are missing, the optional narrative path fails role integrity validation. Duration validation calculates the sum of recommended durations for each candidate narrative node in the optional narrative path and compares this sum with the ad duration in the platform's constraints; if the sum of recommended durations is greater than the ad duration in the platform's constraints, the optional narrative path fails duration validation. Duplicate content validation determines whether there are two or more candidate narrative nodes with the same content tags and belonging to the same narrative role in the optional narrative path; if so, the optional narrative path fails duplicate content validation. Optional narrative paths that pass role integrity validation, duration validation, and duplicate content validation are retained as valid narrative paths.

[0029] When there is only one valid narrative path, that valid narrative path is designated as a candidate narrative path. When there are multiple valid narrative paths, candidate narrative paths are determined according to fixed priority rules. These fixed priority rules include: prioritizing valid narrative paths that contain required nodes; if all paths contain required nodes, prioritizing valid narrative paths whose sum of recommendation durations is closer to and does not exceed the ad duration constraints set by the platform; if recommendation durations are the same, prioritizing valid narrative paths with fewer candidate narrative nodes; and if the number of candidate narrative nodes is the same, prioritizing valid narrative paths with a direct directed transition edge between trusted information nodes and product or service description nodes. Following these fixed priority rules, one or more candidate narrative paths can be determined from multiple valid narrative paths.

[0030] Each candidate narrative path includes sequentially arranged candidate narrative nodes and directed transition edges between adjacent candidate narrative nodes. Candidate narrative nodes correspond to the content units that need to be included in the candidate ad script, and directed transition edges correspond to the connection relationships between adjacent content units. Through the above processing, step S101 obtains the demand state vector, the hierarchical narrative state diagram, and the candidate narrative paths, enabling the advertising needs, brand rules, platform constraints, and historical audience feedback in the short drama ad generation task to be transformed into a clear narrative structure. This improves the structural integrity and content coherence of the short drama ad script in terms of opening attraction, user pain point presentation, product or service description, credible information display, and action guidance.

[0031] Step S102: Call the preset short drama advertisement generation model, generate candidate advertisement scripts based on the demand state vector, hierarchical narrative state diagram and candidate narrative path, and form a candidate advertisement version set based on the candidate advertisement scripts.

[0032] In step S102, the preset short drama advertisement generation model is a pre-trained conditional text generation model, and the model parameters remain fixed during the candidate advertisement script generation process. The preset short drama advertisement generation model includes a conditional input layer, a graph structure encoding layer, a segment state control layer, and a text generation layer. The conditional input layer receives the demand state vector obtained in step S101; the graph structure encoding layer receives the hierarchical narrative state graph obtained in step S101 and encodes the narrative character nodes, candidate narrative nodes, and directed transition edges in the hierarchical narrative state graph to obtain a global graph representation; the segment state control layer receives the candidate narrative path obtained in step S101 and generates the segment state vector corresponding to each script content segment according to the arrangement order of the candidate narrative nodes in the candidate narrative path; the text generation layer generates candidate advertisement scripts under the joint constraints of the demand state vector, the global graph representation, and the segment state vector. Therefore, the preset short drama advertisement generation model is not just a general language model that receives only advertisement demand text, but a script generation model that can receive demand state, narrative graph structure, and narrative path control information.

[0033] Specifically, a script content segment is created for each candidate narrative node in the candidate narrative path, and each script content segment corresponds to a fragment state vector. The fragment state vector is obtained by concatenating narrative role encoding, content tag encoding, recommendation duration encoding, and compliance mark encoding in a fixed order. Specifically, the narrative role encoding is obtained from the category number corresponding to the candidate narrative node's opening attraction, user pain point, product or service description, credible information, or action guidance; the content tag encoding is obtained from the candidate narrative node's content tag encoding; the recommendation duration encoding is obtained from the ratio of the candidate narrative node's recommendation duration to the advertising duration constrained by the platform; and the compliance mark encoding is obtained from the candidate narrative node's corresponding compliance mark encoding. The compliance mark is used to characterize whether the content corresponding to the candidate narrative node involves disclaimers, restrictive expressions, platform-sensitive expressions, or content types requiring compliance prompts. Through the above method, each fragment state vector has a clear data source and calculation method.

[0034] The process of generating candidate ad scripts is represented as follows: ; in, Indicates candidate ad scripts; This represents the demand state vector obtained in step S101; This represents the hierarchical narrative state diagram obtained in step S101. This represents the candidate narrative path obtained in step S101; This indicates the number of generated tags in the candidate ad script; Indicates the sequence number of the generated tag; Indicates the first One generated tag; Indicates the first The text content that was generated before the generation tag; Representation of the graph structure encoding layer to the hierarchical narrative state graph The global graph representation obtained after encoding; Indicates the candidate narrative path The determined first Each generated tag corresponds to a fragment state vector of the script content segment to which it belongs; This indicates that, under the combined constraints of the already generated text content, the demand state vector, the global graph representation, and the fragment state vector, the text generation layer generates the first... The conditional probability of each generated tag. Candidate narrative paths. Used to determine the generation order of each script content segment, and to determine the first... The fragment state vector corresponding to each generated tag .

[0035] When generating candidate ad scripts, the text generation layer generates each script content segment sequentially according to the order of the candidate narrative nodes in the candidate narrative path. The opening attraction node corresponds to the ad's intro content; the user pain point node corresponds to the target audience's needs or usage scenario issues; the product or service description node corresponds to the content linking the product or service to the user's pain point; the credible information node corresponds to evidence, reviews, case studies, or guarantee information; and the action guidance node corresponds to purchase, consultation, click, lead generation, or redirect content. These script content segments are combined in the order of the candidate narrative path to obtain the candidate ad script.

[0036] After obtaining the candidate ad scripts, basic verification is performed. This includes content unit integrity verification, brand rule verification, platform rule verification, and script duration verification. Content unit integrity verification determines whether the candidate ad script contains script content segments corresponding to opening attraction, user pain points, product or service descriptions, credible information, and action guidance. Brand rule verification determines whether the candidate ad script contains content that violates brand tone, prohibited expressions, disclaimer rules, or compliant expression requirements. Platform rule verification determines whether the candidate ad script violates the platform's review rules. Script duration verification determines whether the candidate ad script exceeds the ad duration constrained by the platform based on the sum of the recommended durations of each script content segment. Candidate ad scripts that pass the basic verification are retained, while those that fail are removed. When a removed candidate ad script needs to be regenerated, it is regenerated according to the corresponding candidate narrative path, or a new candidate ad script is generated according to the candidate narrative path ranked later in step S101.

[0037] After retaining the candidate ad scripts, a candidate ad version set is formed based on them. Specifically, shot instructions, narration text, and subtitle schemes are generated according to the narrative roles and content tags of each script content segment in the candidate ad scripts. Shot instructions include shot content, shot order, and shot duration; narration text is derived from spoken or narrative content in the candidate ad scripts; and subtitle schemes are derived from key expressions in the candidate ad scripts and subtitle area requirements in the platform constraints. When the task association information includes existing material indexes, material call information is generated based on the shot instructions and existing material indexes. Material call information includes the material type, material tag, and material storage location matching the shot instructions. The script text, shot instructions, narration text, subtitle scheme, and material call information corresponding to the same candidate ad script are associated as one candidate ad version, and multiple candidate ad versions are aggregated into a candidate ad version set. When the task association information does not include existing material indexes, the candidate ad version does not generate material call information, or the material call information is marked as empty.

[0038] Through the above processing, step S102 yields a set of candidate ad scripts and candidate ad versions. The candidate ad scripts are constrained by the demand state vector, hierarchical narrative state diagram, and candidate narrative paths during generation, ensuring that the script content is generated according to a defined narrative structure. The set of candidate ad versions is formed based on the candidate ad scripts and retains the correspondence between the candidate ad scripts and candidate narrative paths, allowing feedback data collected during the deployment test to be traced back to the corresponding candidate ad script and its narrative structure.

[0039] Step S103: Conduct delivery tests on the candidate ad versions in the candidate ad version set and collect the original feedback data corresponding to each candidate ad version; perform feedback attribution and noise reduction processing on the original feedback data to obtain the attribution feedback results corresponding to each candidate ad version.

[0040] In step S103, a small-scale test is conducted on each candidate ad version in the candidate ad version set obtained in step S102, and the raw feedback data corresponding to each candidate ad version is collected. The raw feedback data comes from the ad delivery logs and performance statistics logs returned by the ad delivery platform to the advertiser's account. The raw feedback data includes candidate ad version identifier, exposure results, click results, conversion results, viewing duration, transaction amount, delivery time, exposure position, budget tier, audience segmentation, and delivery platform. The above raw feedback data uses candidate ad versions, exposure samples, or delivery batches as statistical objects and does not include personal information that can directly identify natural persons; when it is necessary to identify duplicate trigger records, the anonymous session identifier or de-identified trigger identifier returned by the ad delivery platform is used for judgment. By collecting feedback indicators and delivery context simultaneously, the performance of ad content and the impact of the delivery environment can be distinguished when calculating the ad version feedback results, avoiding the direct use of raw click-through rate, conversion rate, or transaction data as the basis for model optimization.

[0041] Specifically, during the campaign testing process, each candidate ad version in the candidate ad version set is assigned a version identifier, and feedback records are generated based on the exposure logs, click logs, conversion logs, viewing logs, and transaction logs recorded by the campaign platform. Each feedback record corresponds to one exposure sample and includes at least the candidate ad version identifier, exposure result, whether there was a click, whether there was a conversion, viewing duration, transaction amount, and the corresponding campaign context; the campaign context consists of the campaign time period, exposure position, budget tier, audience segment, and campaign platform. The feedback records are aggregated according to the candidate ad version identifier to obtain the raw feedback data corresponding to each candidate ad version.

[0042] After collecting the raw feedback data, noise reduction processing is performed. This includes invalid record removal and abnormal record removal. Invalid record removal refers to removing feedback records that lack candidate ad version identifiers, lack delivery context, lack exposure results, or fail to transmit transaction amount. Abnormal record removal refers to removing feedback records where the same anonymous session identifier or de-identified trigger identifier repeatedly triggers clicks on the same candidate ad version more than a preset number of times within a preset time window, as well as feedback records with zero viewing time, which cannot match a valid exposure path but still have conversion results. The feedback records retained after noise reduction are considered valid feedback samples. Through these processes, the interference of missing logs, duplicate triggers, and abnormal transmissions on the feedback calculation results can be reduced.

[0043] After obtaining valid feedback samples, first calculate the feedback benefit observation for each valid feedback sample. For candidate ad versions... The For each valid feedback sample, the observed feedback benefit value is determined according to the following formula: ; in, Indicates candidate ad version The Feedback benefit observations for each valid feedback sample; Indicates a click mark, at the... A valid feedback sample is assigned a value of 1 if a click occurs, and a value of 0 if no click occurs. Indicates the conversion marker, in the... The value is 1 when a valid feedback sample is converted, and 0 when no conversion occurs. Indicates the percentage of viewers who have completed watching. This is the ratio of the actual viewing time to the duration of the candidate ad version; when the ratio is greater than 1, it is set to 1. This represents the standardized value of the transaction amount. For the first The ratio of the transaction amount corresponding to each valid feedback sample to the preset transaction amount benchmark value; when the ratio is greater than 1, it is taken as 1. Indicates click weight. Indicates conversion weight. Indicates the viewing weight. Indicates the weight of transactions. , , and All four are preset non-negative weights, and their sum is 1. In this embodiment, the click marker, conversion marker, viewing completion rate, and standardized transaction amount have the same influence ratio in the feedback revenue observation. , , and All values ​​are fixed weights, each with a value of 0.25. The preset transaction amount benchmark is the normalized benchmark for transaction amount set by the advertiser in the campaign testing task, and the preset transaction amount benchmark value is greater than 0.

[0044] After calculating the observed feedback revenue, feedback attribution processing is performed on the valid feedback samples. First, valid feedback samples are grouped contextually according to the delivery time, exposure location, budget tier, audience segment, and delivery platform. Valid feedback samples within the same contextual group share the same delivery time, exposure location, budget tier, audience segment, and delivery platform. For any valid feedback sample, its baseline revenue is the average of the observed feedback revenue of all valid feedback samples within its contextual group. If the number of valid feedback samples in a contextual group is lower than a preset minimum sample size, the valid feedback samples in that contextual group are merged into adjacent delivery time groups under the same delivery platform and audience segment for baseline revenue calculation. If the merged sample size is still lower than the preset minimum sample size, the valid feedback samples in that contextual group are not included in the calculation of the attribution feedback results for this round. Therefore, the baseline revenue is calculated only from valid feedback samples with similar delivery contexts in this delivery test, and data from uncertain sources is not used.

[0045] Candidate Ad Versions The causal correction feedback value in the attribution feedback results is expressed as: ; in, Indicates candidate ad version The causal correction feedback value; Indicates candidate ad version The number of valid feedback samples that participate in the calculation of attribution feedback results after denoising processing, and Greater than 0; Indicates candidate ad version The corresponding number One valid feedback sample; Indicates candidate ad version The Feedback benefit observations for each valid feedback sample; Indicates the first The delivery context corresponding to each valid feedback sample; Indicates the context of the delivery. The corresponding baseline revenue, which is the revenue generated in this test campaign within the context of the campaign. The average of the observed feedback returns for all valid feedback samples belonging to the same context group; Indicates the candidate ad version to be displayed. The delivery action; Indicates the context of the delivery. Next candidate ad version The estimated propensity to be assigned a delivery is the delivery context in this delivery test. Candidate ad versions within the corresponding context group The ratio between the number of times an ad is displayed and the total number of times all candidate ad versions within that context group are displayed; This represents a preset stability constant, with a value greater than 0 and less than 1. In this embodiment, because... The purpose is simply to prevent the denominator from being too small due to an excessively low probability of bias; therefore, a fixed value of 0.01 is the most stable. This indicates that the larger value is chosen between the estimated propensity and the preset stability constant.

[0046] Based on the above calculations Used to deduct the basic feedback impact caused by the combined effects of the time period, exposure location, budget tier, audience segmentation, and advertising platform within the same advertising context; Used to characterize candidate ad versions The proportion of exposure opportunities obtained under the same delivery context; by estimating the propensity to adjust, the impact of differences in traffic allocation opportunities among different candidate ad versions on the feedback judgment can be reduced. The causal adjustment feedback value can be positive, zero, or negative. A positive value indicates that the candidate ad version performs better than the baseline after deducting the influence of the baseline under the same delivery context and adjusting for differences in exposure allocation, while a negative value indicates that the candidate ad version performs worse than the baseline after the above adjustment.

[0047] After calculating the causal correction feedback value of the candidate ad version, the candidate ad version identifier, the number of valid feedback samples, the denoised click statistics, the denoised conversion statistics, the denoised view statistics, the denoised transaction statistics, and the causal correction feedback value are combined to obtain the attribution feedback result corresponding to the candidate ad version. Through the above processing, step S103 no longer directly judges the quality of the ad version based on the original feedback indicators, but first removes invalid and abnormal feedback records, and then calculates the baseline revenue and estimated propensity based on the same delivery context, and performs attribution correction on the original feedback data, thereby improving the reliability of delivery feedback when used for model optimization.

[0048] Step S104: Construct preference samples among candidate ad versions based on attribution feedback results, and optimize and update the preset short drama ad generation model based on the preference samples and the narrative structure constraints corresponding to the hierarchical narrative state diagram.

[0049] In step S104, based on the attribution feedback results corresponding to each candidate ad version obtained in step S103, a preference sample is constructed among the candidate ad versions. Then, the candidate ad scripts in the preference sample are mapped to the hierarchical narrative state diagram obtained in step S101, and the structural advantage difference between the preferred script and the inferior script is calculated. Finally, the preference sample and the structural advantage difference are jointly introduced into the preference optimization loss to optimize and update the preset short drama ad generation model. Through this process, the model update is not directly based on the original click-through rate, conversion rate, or transaction data, but rather on the feedback results after attribution correction and the narrative structure constraints corresponding to the hierarchical narrative state diagram. This allows the short drama ad generation model to learn the preference relationships of the delivery feedback while maintaining the integrity and sequential consistency of the script content units.

[0050] Specifically, preference samples are first constructed based on the attribution feedback results corresponding to each candidate ad version. For any two candidate ad versions... and Read candidate ad versions Causal correction feedback value and the number of valid feedback samples used in the calculation of attribution feedback results and candidate ad versions Causal correction feedback value and the number of valid feedback samples used in the calculation of attribution feedback results .when and All are not less than the preset minimum sample size, and When the difference exceeds the preset feedback threshold, the candidate ad version will be... The corresponding candidate ad script is selected as the preferred script, and the candidate ad version is... The corresponding candidate ad scripts are used as the inferior ad scripts, and the candidate ad scripts corresponding to the two are combined into a preference sample; when and All are not less than the preset minimum sample size, and When the difference exceeds the preset feedback threshold, the candidate ad version will be... The corresponding candidate ad script is selected as the preferred script, and the candidate ad version is... The corresponding candidate ad scripts are used as inferior scripts, and the candidate ad scripts corresponding to the two are combined into a preference sample. When the number of effective feedback samples of the two candidate ad versions does not reach the preset minimum sample size, or the difference between their causal correction feedback values ​​does not exceed the preset feedback difference threshold, the two candidate ad versions are not used to construct the preference sample. This yields the preference sample set. Preference sample set Each preference sample in the dataset includes both a preferred script and a disliked script.

[0051] After obtaining the preference sample set, calculate the structural quality scores of the preferred and unpredictable scripts relative to the hierarchical narrative state graph within the preference samples. For any candidate ad script... First, select the candidate ad scripts. The script content segments are mapped to the candidate narrative nodes in their corresponding candidate narrative paths, and then the node coverage ratio, sequence consistency ratio, and rule satisfaction ratio are calculated respectively. The node coverage ratio is the candidate ad script The ratio of the number of essential narrative characters matching the corresponding candidate narrative path to the total number of essential narrative characters. Essential narrative characters include opening grabs, user pain points, product or service descriptions, credible information, and action guidance; the order consistency ratio is based on the candidate ad scripts. The number of candidate ad scripts whose adjacent script content segments are in the same order as the adjacent candidate narrative nodes in the corresponding candidate narrative path. The ratio of the total number of adjacent script content segments in a candidate ad script; the rule satisfies the ratio. The number of script content segments that did not violate brand rules and the platform's review rules accounted for the majority of candidate ad scripts. The ratio of the total number of script content segments. The average of the node coverage ratio, the order consistency ratio, and the rule fulfillment ratio is used to determine the candidate ad script. Structural mass fraction Among them, when the candidate ad script When there are no adjacent script content segments, the consistency ratio is 0; when the candidate ad script When no script content segment exists, the node coverage ratio, sequence consistency ratio, and rule satisfaction ratio are all set to 0. This limitation prevents the structure quality score from being uncalculated due to insufficient script content segments.

[0052] For the preferred script in the preference sample and inferior selection script The structural advantage difference is expressed as: ; in, Indicates the preferred script Compared to the inferior selection script Poor structural advantages; Indicates the preferred script The structural mass fraction; Indicates the script for the inferior selection The structural quality score. By using this structural advantage difference, the node coverage, node order, and rule satisfaction of the hierarchical narrative state diagram can be introduced into the model optimization process, avoiding the weakening of the script narrative structure by the model learning solely based on the feedback preference.

[0053] After obtaining the preference sample set and structural advantage difference, the preset short drama advertisement generation model is optimized and updated. The preset short drama advertisement generation model from step S102 is used as the current model to be optimized, and the model parameters are denoted as... The original short drama ad generation model was used as the reference model, and the parameters of the reference model are denoted as follows: And refer to model parameters This remains constant during this round of optimization. For preference samples... The current model to be optimized is related to the preferred script. and inferior selection script The logarithmic difference of the generation probability is denoted as , This is equivalent to generating the optimal script for the current model to be optimized. The log probability minus the generated inferior selection script Logarithmic probability; reference model for optimal script and inferior selection script The logarithmic difference of the generation probability is denoted as , Equals the reference model to generate the preferred script The log probability minus the generated inferior selection script The logarithmic probability of the script is obtained by multiplying the conditional probabilities of each generated label in the script by the corresponding model according to the conditional generation process in step S102. The logarithmic probability of the script is the sum of the logarithmic values ​​of the conditional probabilities of each generated label.

[0054] The preference optimization loss is expressed as: ; in, Indicate the loss of preference optimization; Represents a set of preference samples; Represents the set of preference samples A preference sample in which candidate ad versions Corresponding preferred script Candidate Ad Versions Corresponding inferior selection script ; Represents a logarithmic function; Represents the Sigmoid function; This indicates the current model to be optimized relative to the preferred script. and inferior selection script The logarithmic difference of the generation probability; Indicates the reference model for the preferred script and inferior selection script The log difference of generation probabilities; both the current model to be optimized and the reference model follow the candidate advertisement script generation process in step S102 to calculate the conditional probability of each generated marker in the preferred script and the inferior script respectively; sum the log values ​​of the conditional probabilities of each generated marker in the same script to obtain the log value of the generation probability corresponding to the script; then subtract the log value of the generation probability corresponding to the inferior script from the log value of the generation probability corresponding to the preferred script to obtain the log difference of generation probabilities between the preferred script and the inferior script under the corresponding model; Indicates the preferred script Compared to the inferior selection script Poor structural advantages; Represents the structural constraint weights. To pre-determine non-negative weights, the influence of structural advantage difference on preference optimization loss is controlled. Specifically, the values ​​of the structural constraint weights are determined before model optimization and remain unchanged during this round of model optimization. The preference sample set is divided into a training preference sample set and a validation preference sample set. Candidate structural constraint weight sets are generated within a value range greater than 0 and less than 1 with a fixed step size of 0.1. The candidate structural constraint weight set is... Each candidate structural constraint weight from the candidate structural constraint weight set is used to pre-update the current model to be optimized for the same number of rounds. The preference discrimination accuracy and average structural quality score of the pre-updated model are then calculated on the validation preference sample set. Candidate structural constraint weights whose average structural quality score after pre-update is not lower than the original average structural quality score are selected as optional structural constraint weights. From these optional structural constraint weights, the candidate structural constraint weight with the highest preference discrimination accuracy is selected as the structural constraint weight used for this round of model optimization and update. When multiple optional structural constraint weights have the same preference discrimination accuracy, the candidate structural constraint weight with the smaller value is selected.

[0055] Wherein, the original average structure quality score represents the average structure quality score of the scripts generated by the current model to be optimized on the set of validation preference samples before the pre-update; the average structure quality score of the model after the pre-update represents the average structure quality score of the scripts generated on the set of validation preference samples after the pre-update is completed using candidate structure constraint weights; the preference discrimination accuracy represents the proportion of the number of preference samples in the set of validation preference samples where the logarithm of the generation probability of the preferred script is higher than the logarithm of the generation probability of the corresponding undesired script.

[0056] When optimizing and updating the model, minimize the preference optimization loss. To achieve this, the trainable parameters in the current model to be optimized are updated. Since the preference samples are selected by combining the difference in causal correction feedback values ​​and the number of effective feedback samples, the preference relationships involved in the update have excluded version combinations with insufficient sample size and insignificant feedback differences; and since the structural advantage difference is introduced into the preference optimization loss... The model update is simultaneously influenced by attribution feedback preference relationships and narrative structure constraints. Through the above processing, step S104 yields the optimized and updated short drama ad generation model, making the model more inclined to generate short drama ad scripts with better attribution feedback performance and in line with the constraints of the hierarchical narrative state diagram.

[0057] In summary, this method constructs a demand state vector by acquiring task association information for short drama ad generation, and generates a hierarchical narrative state diagram based on this vector to describe ad narrative units and their transition relationships. This determines candidate narrative paths, ensuring that short drama ad scripts form structured constraints on ad narrative units and their sequential connections before generation. Based on this, a pre-defined short drama ad generation model is invoked to generate candidate ad scripts based on the demand state vector, hierarchical narrative state diagram, and candidate narrative paths, forming a set of candidate ad versions. This allows candidate ad content to be organized according to a clear narrative path. Subsequently, the candidate ad versions are tested and raw feedback data is collected. Attribution feedback results are obtained through feedback attribution and noise reduction, ensuring that the feedback is no longer directly based on raw feedback data but reflects the feedback performance of candidate ad versions after attribution and noise reduction. Finally, preference samples among candidate ad versions are constructed based on the attribution feedback results, and the pre-defined short drama ad generation model is optimized and updated in conjunction with the narrative structure constraints corresponding to the hierarchical narrative state diagram. Therefore, this application, on the one hand, uses hierarchical narrative state diagrams and candidate narrative paths to structurally control the content units and their connections in the short drama advertisement script; on the other hand, it improves the reliability of the delivery feedback when participating in model optimization through attribution feedback results and preference samples. Thus, it can take into account both script narrative structure control and reliable utilization of delivery feedback in the process of optimizing the short drama advertisement generation model, so that the updated generation model can more stably output short drama advertisement content that is structurally complete, content-coherent, and more in line with the delivery effect requirements.

[0058] Figure 2 This is a structural block diagram of a short drama advertisement generation model optimization system based on feedback attribution provided in one embodiment of this application. The system includes at least the following modules: The narrative path construction module is used to obtain the task association information of the short drama advertisement generation task, construct the demand state vector according to the task association information, generate a hierarchical narrative state diagram to describe the advertisement narrative unit and its transition relationship based on the demand state vector, and determine the candidate narrative path corresponding to the hierarchical narrative state diagram. The ad version generation module is used to call a preset short drama ad generation model, generate candidate ad scripts based on the demand state vector, the hierarchical narrative state diagram and the candidate narrative path, and form a candidate ad version set based on the candidate ad scripts; The feedback attribution processing module is used to conduct delivery tests on candidate ad versions in the candidate ad version set, collect raw feedback data corresponding to each candidate ad version, and perform feedback attribution and noise reduction processing on the raw feedback data to obtain the attribution feedback results corresponding to each candidate ad version. The model optimization and update module is used to construct preference samples among the candidate ad versions based on the attribution feedback results, and to optimize and update the preset short drama ad generation model based on the preference samples and the narrative structure constraints corresponding to the hierarchical narrative state diagram.

[0059] For relevant details, please refer to the above method implementation examples.

[0060] Figure 3 This is a block diagram of an electronic device provided in one embodiment of this application. The device includes at least a processor 301 and a memory 302.

[0061] Processor 301 may include one or more processing cores, such as a quad-core processor or an octa-core processor. Processor 301 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 301 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 301 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content required to be displayed on the screen. In some embodiments, processor 301 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0062] Memory 302 may include one or more computer-readable storage media, which may be non-transitory. Memory 302 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in memory 302 is used to store at least one instruction, which is executed by processor 301 to implement the feedback attribution-based short drama advertisement generation model optimization method provided in the method embodiments of this application.

[0063] In some embodiments, the electronic device may optionally include a peripheral device interface and at least one peripheral device. The processor 301, memory 302, and peripheral device interface can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface via a bus, signal line, or circuit board. Indicatively, peripheral devices include, but are not limited to, radio frequency circuits, touch displays, audio circuits, and power supplies.

[0064] Of course, electronic devices may also include fewer or more components, and this embodiment does not limit this.

[0065] Optionally, this application also provides a computer-readable storage medium storing a program that is loaded and executed by a processor to implement the feedback attribution-based short drama advertising generation model optimization method of the above method embodiments.

[0066] Optionally, this application also provides a computer product including a computer-readable storage medium storing a program that is loaded and executed by a processor to implement the feedback attribution-based short drama advertising generation model optimization method of the above method embodiments.

[0067] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0068] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for optimizing a short drama advertisement generation model based on feedback attribution, characterized in that, The method includes: Obtain task association information for short drama ad generation tasks, construct a demand state vector based on the task association information, and generate a hierarchical narrative state diagram based on the demand state vector to describe the ad narrative unit and its transition relationship, and determine the candidate narrative path corresponding to the hierarchical narrative state diagram. A preset short drama ad generation model is invoked to generate candidate ad scripts based on the demand state vector, the hierarchical narrative state diagram, and the candidate narrative path, and a candidate ad version set is formed based on the candidate ad scripts; Test the candidate ad versions in the candidate ad version set and collect the original feedback data corresponding to each candidate ad version; perform feedback attribution and noise reduction processing on the original feedback data to obtain the attribution feedback results corresponding to each candidate ad version. Based on the attribution feedback results, a preference sample is constructed among the candidate ad versions, and the preset short drama ad generation model is optimized and updated based on the preference sample and the narrative structure constraints corresponding to the hierarchical narrative state diagram.

2. The method for optimizing the short drama advertisement generation model based on feedback attribution according to claim 1, characterized in that, The step of generating a hierarchical narrative state diagram based on the demand state vector to describe advertising narrative units and their transition relationships, and determining candidate narrative paths corresponding to the hierarchical narrative state diagram, includes: A hierarchical narrative state diagram is generated based on a preset narrative unit library, which includes narrative character nodes and fragment instance nodes, wherein the fragment instance nodes are specific content fragments under each narrative character node. Based on the product category, target audience, brand rules, and platform constraints determined in the task association information, and in conjunction with the demand state vector, the fragment instance nodes in the preset narrative unit library are filtered, the filtered fragment instance nodes are determined as candidate narrative nodes, and the candidate narrative nodes are connected to their respective narrative role nodes. When the relationship between the narrative roles of two candidate narrative nodes conforms to the preset narrative order rules, and there are no conflicts in content tags, brand rules, or platform constraints between them, a directed transfer edge is established between the two candidate narrative nodes. From the hierarchical narrative state diagram, a candidate narrative path is determined with the candidate narrative node corresponding to the opening attraction node as the path start point, the candidate narrative node corresponding to the action guidance node as the path end point, and the candidate narrative nodes corresponding to the user pain point node, product or service description node and credible information node are sequentially included between the path start point and the path end point.

3. The method for optimizing the short drama advertisement generation model based on feedback attribution according to claim 1, characterized in that, The process involves calling a preset short drama ad generation model, generating candidate ad scripts based on the demand state vector, the hierarchical narrative state diagram, and the candidate narrative path, and forming a candidate ad version set based on the candidate ad scripts, including: A pre-trained conditional text generation model is invoked as the preset short drama advertisement generation model. The conditional text generation model includes a conditional input layer, a graph structure encoding layer, a segment state control layer, and a text generation layer. The conditional input layer receives the demand state vector, and the graph structure encoding layer encodes the narrative character nodes, candidate narrative nodes, and directed transition edges in the hierarchical narrative state graph to obtain a global graph representation. The fragment state control layer generates the fragment state vector corresponding to each script content segment according to the arrangement order of the candidate narrative nodes in the candidate narrative path. Under the joint constraints of the demand state vector, the global graph representation, and the fragment state vector, the text generation layer generates each script content segment sequentially according to the arrangement order of the candidate narrative nodes in the candidate narrative path, and combines each script content segment to obtain the candidate advertisement script. The candidate ad scripts are verified for content unit integrity, brand rules, platform rules, and script duration. Candidate ad scripts that pass the verification are retained. Based on the retained candidate ad scripts, camera instructions, narration text, and subtitle schemes are generated. The script text, camera instructions, narration text, and subtitle schemes corresponding to the same candidate ad script are associated as candidate ad versions. Multiple candidate ad versions are aggregated to obtain a candidate ad version set.

4. The method for optimizing the short drama advertisement generation model based on feedback attribution according to claim 1, characterized in that, The step involves testing the delivery of candidate ad versions from the candidate ad version set and collecting raw feedback data for each candidate ad version, including: Assign a version identifier to each candidate ad version in the candidate ad version set, and generate feedback records based on the exposure logs, click logs, conversion logs, viewing logs, and transaction logs recorded by the ad delivery platform; Each feedback record corresponds to one exposure sample, and each feedback record includes the candidate ad version identifier, whether it was clicked, whether it was converted, the viewing duration, the transaction amount, and the corresponding placement context. The placement context includes the placement time period, the exposure position, the budget level, the audience segment, and the placement platform. The feedback records were aggregated according to the candidate ad version identifier to obtain the original feedback data corresponding to each candidate ad version; The original feedback data is subjected to invalid record removal and abnormal record removal. The feedback records retained after noise reduction are used as valid feedback samples. The invalid record removal includes deleting feedback records that lack candidate ad version identifiers, lack delivery context, lack exposure results, or fail to transmit transaction amount. The abnormal record removal includes deleting feedback records that repeatedly trigger clicks on the same candidate ad version more than a preset number of times within a preset time window using the same anonymous session identifier or de-identified trigger identifier, as well as feedback records with zero viewing time, which cannot match the effective exposure link but have conversion results.

5. The method for optimizing the short drama advertisement generation model based on feedback attribution according to claim 4, characterized in that, The process of performing feedback attribution and denoising on the original feedback data to obtain the attribution feedback results corresponding to each candidate ad version includes: Based on the standardized values ​​of click markers, conversion markers, view completion rate, and transaction amount corresponding to the valid feedback samples, calculate the feedback revenue observation value of the valid feedback samples; The valid feedback samples were grouped into contexts based on the time period, exposure location, budget level, audience segmentation, and advertising platform. The average of the feedback revenue observations of all valid feedback samples within the same context group was determined as the baseline revenue for the corresponding advertising context. The probability of a candidate ad version being assigned to a specific delivery context is determined by the ratio of the number of times a candidate ad version is exposed within the same context group to the total number of times all candidate ad versions within that context group are exposed. Based on the observed feedback revenue, the baseline revenue, and the propensity probability, a causal correction feedback value for the candidate ad version is calculated, and the causal correction feedback value is expressed as follows: ； in, Indicates candidate ad version The causal correction feedback value; Indicates candidate ad version The number of valid feedback samples that participate in the calculation of attribution feedback results after denoising processing, and Greater than 0; Indicates candidate ad version The corresponding number One valid feedback sample; Indicates candidate ad version The Feedback benefit observations for each valid feedback sample; Indicates the first The delivery context corresponding to each valid feedback sample; Indicates the context of the delivery. Corresponding baseline returns; Indicates the candidate ad version to be displayed. The delivery action; Indicates the context of the delivery. Next candidate ad version The probability of being assigned a distribution; This represents the preset stability constant; The candidate ad version identifier, the number of valid feedback samples, the denoised click statistics, the denoised conversion statistics, the denoised view statistics, the denoised transaction statistics, and the causal correction feedback value are combined to obtain the attribution feedback result corresponding to the candidate ad version.

6. The optimization method for short drama advertisement generation model based on feedback attribution according to claim 1, characterized in that, The step of constructing preference samples among the candidate ad versions based on the attribution feedback results includes: For any two candidate ad versions, read the causal correction feedback value and the number of valid feedback samples corresponding to the two candidate ad versions respectively; When the number of valid feedback samples for both candidate ad versions is not less than the preset minimum number of samples, and the difference between the causal correction feedback value of one candidate ad version and the causal correction feedback value of the other candidate ad version is greater than the preset feedback difference threshold, the candidate ad version with the higher causal correction feedback value is selected as the preferred version, and the candidate ad version with the lower causal correction feedback value is selected as the inferior version. The candidate ad scripts corresponding to the preferred version and the inferior version are respectively used to form a preference sample. If the number of valid feedback samples for two candidate ad versions does not reach the preset minimum number of samples, or if the difference between the causal correction feedback values ​​of the two candidate ad versions does not exceed the preset feedback difference threshold, then the two candidate ad versions will not be used to construct preference samples. A preference sample set is formed based on the constructed preference samples.

7. The method for optimizing the short drama advertisement generation model based on feedback attribution according to claim 6, characterized in that, The optimization and update of the preset short drama advertisement generation model based on the preference samples and the narrative structure constraints corresponding to the hierarchical narrative state diagram includes: The candidate ad scripts in the preference sample are mapped to the candidate narrative nodes in the corresponding candidate narrative path, and the structural quality score of the candidate ad scripts is determined based on the average of the node coverage ratio, the order consistency ratio, and the rule satisfaction ratio. The structural dominance difference is determined by the difference between the structural quality scores of the preferred scripts and the undesired scripts in the preference sample. ； in, Indicates the preferred script Compared to the inferior selection script Poor structural advantages; Indicates the preferred script The structural mass fraction; Indicates the script for the inferior selection The structural mass fraction; Based on the log difference in generation probabilities of the current model to be optimized for the preferred and undesired scripts, the log difference in generation probabilities of the reference model for the preferred and undesired scripts, and the aforementioned structural advantage difference, a preference optimization loss is constructed: ； in, Indicates preference optimization loss; Represents a set of preference samples; Represents the set of preference samples A preference sample in which candidate ad versions Corresponding preferred script Candidate Ad Versions Corresponding inferior selection script ; Represents a logarithmic function; Represents the Sigmoid function; This indicates the current model to be optimized relative to the preferred script. and inferior selection script The logarithmic difference of the generation probability; Indicates the reference model for the preferred script and inferior selection script The logarithmic difference of the generation probability; Indicates the preferred script Compared to the inferior selection script Poor structural advantages; Represents the structural constraint weights. Preset non-negative weights; With the goal of minimizing the preference optimization loss, the trainable parameters in the preset short drama advertisement generation model are updated.

8. A short drama advertisement generation model optimization system based on feedback attribution, characterized in that, include: The narrative path construction module is used to obtain the task association information of the short drama advertisement generation task, construct the demand state vector according to the task association information, generate a hierarchical narrative state diagram to describe the advertisement narrative unit and its transition relationship based on the demand state vector, and determine the candidate narrative path corresponding to the hierarchical narrative state diagram. The ad version generation module is used to call a preset short drama ad generation model, generate candidate ad scripts based on the demand state vector, the hierarchical narrative state diagram and the candidate narrative path, and form a candidate ad version set based on the candidate ad scripts; The feedback attribution processing module is used to conduct delivery tests on candidate ad versions in the candidate ad version set, collect raw feedback data corresponding to each candidate ad version, and perform feedback attribution and noise reduction processing on the raw feedback data to obtain the attribution feedback results corresponding to each candidate ad version. The model optimization and update module is used to construct preference samples among the candidate ad versions based on the attribution feedback results, and to optimize and update the preset short drama ad generation model based on the preference samples and the narrative structure constraints corresponding to the hierarchical narrative state diagram.

9. An electronic device, characterized in that, The device includes a processor and a memory; the memory stores a program that is loaded and executed by the processor to implement a feedback attribution-based short drama advertising generation model optimization method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program that, when executed by a processor, is used to implement a short drama advertisement generation model optimization method based on feedback attribution as described in any one of claims 1 to 7.

Images