Advertisement material pre-auditing method and system based on e-commerce putting scene

Abstract

The invention discloses an advertisement material pre-auditing method and system based on an e-commerce putting scene, and relates to the technical field of advertisements. The method comprises the steps of obtaining advertisement materials and putting context information, and performing multi-modal analysis to generate a pre-auditing data packet; retrieving related advertisement specifications based on a rule knowledge base corresponding to the advertisement specifications of the e-commerce platform; performing compliance detection on the pre-auditing data packet by using a first multi-modal large language model under a calling prompt containing the specification as a context to generate a compliance conclusion; based on the material internal characteristics and the delivery context information in the pre-audited data packet, performing delivery suitability evaluation by using a pre-trained click rate prediction model, and generating an suitability score; and integrating the compliance conclusion and the suitability score to generate a pre-auditing report.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of advertising technology, and in particular to an advertising material pre-audit method and system based on e-commerce placement scenarios. BACKGROUND

[0002] With the rapid development of e-commerce, online advertising placement has become the core link of business marketing. As the key carrier connecting goods and consumers, the compliance and placement effect of advertising materials directly determine the success or failure and cost-effectiveness of marketing activities.

[0003] Currently, the review of advertising materials mainly relies on the "post-placement review" mechanism of e-commerce platforms after the advertising principal submits a placement plan. This mode has significant drawbacks: once the material is rejected due to violation of platform specifications, not only will the advertising plan be delayed online, affecting the marketing rhythm, but also the advertising account weight may be reduced due to the violation record, which will have a negative impact on subsequent placement. To avoid this risk, some advertisers use manual "pre-placement pre-audit", that is, manually check the complex and dynamically updated advertising rules published by the platform. However, this method is inefficient, time-consuming and labor-intensive, and is prone to errors when faced with a large amount of materials.

[0004] Although some existing automatic review tools have improved efficiency, they mostly focus on compliance detection, and their built-in rule base or model often cannot quickly respond to the frequent updates of the review policy of e-commerce platforms, resulting in deviations between the pre-audit results and the final review standards of the platform. More critically, both manual pre-audit and existing automatic tools generally lack the ability to predictively assess the placement effect of advertising materials in specific placement scenarios (such as specific ad positions, target audience). A compliant material is not equivalent to an effective material with a high Click-Through Rate (CTR), and advertisers still face the risk of investing a large budget in compliant but inefficient materials, resulting in waste of advertising resources.

[0005] Therefore, the industry urgently needs a means that can accurately and dynamically evaluate the compliance of advertising materials before placement, and scientifically predict their placement adaptability and potential effect. SUMMARY

[0006] The purpose of the present application is to solve the problems in the prior art, and an advertising material pre-audit method and system based on e-commerce placement scenarios are proposed. In order to achieve the above purpose, the technical solutions adopted by the embodiments of the present application are as follows:

[0007] In a first aspect, the embodiments of the present application propose an advertising material pre-audit method based on e-commerce placement scenarios, comprising the following steps:

[0008] The system acquires the advertising creatives to be pre-approved and the placement context information associated with the advertising creatives, and performs multimodal parsing on the advertising creatives to generate a pre-approval data package. The pre-approval data package contains structured visual features, text features and auditory features extracted from the advertising creatives.

[0009] Based on a pre-built rule knowledge base that dynamically corresponds to the advertising specifications of e-commerce platforms, relevant advertising specifications that match the context information of the campaign are retrieved.

[0010] The first multimodal large language model is used to perform compliance checks on the pre-screening data packets. The compliance checks are based on call prompts that include relevant advertising specifications as context, in order to generate preliminary compliance conclusions.

[0011] The inherent features of the creative materials are extracted from the pre-screening data package. The inherent features of the creative materials are at least a part of the structured visual features, text features and auditory features. Based on the inherent features of the creative materials and the context information of the delivery, a pre-trained click-through rate prediction model is used to evaluate the delivery suitability of the advertising creative materials in order to generate a predicted suitability score.

[0012] In addition, the preliminary compliance conclusions and suitability scores are integrated to generate a pre-audit report that includes compliance status, suitability score, and optimization suggestions.

[0013] In a possible design, the steps for multimodal analysis of advertising creatives include:

[0014] If the advertising material is a video, then automatic speech recognition is used to extract the text of the spoken words corresponding to the spoken audio in the video material;

[0015] In addition, computer vision is used to extract visual elements from advertising materials. Visual elements include at least: brand logo, text extracted through optical character recognition, and the main body of the product.

[0016] In one possible design, the rule knowledge base is constructed using a hierarchical classification system, which divides advertising norms into multiple levels from domain, theme, subtype to behavior, in order to achieve path-based risk identification.

[0017] In one possible design, after performing compliance checks, the following is also included:

[0018] Determine whether the pre-screening data packet contains a statement to be verified that requires external information.

[0019] If included, at least one verification action is dynamically planned and executed to retrieve verification evidence from external data sources and update the preliminary compliance conclusions based on the verification evidence; verification actions include reverse image search or web search.

[0020] In a possible design, after performing the compliance detection, the method further includes:

[0021] If the output confidence of the first multi-modal large language model on the preliminary compliance conclusion is lower than a preset confidence threshold, the pre-audit data package and the preliminary compliance conclusion are submitted to a second multi-modal large language model as a parent model;

[0022] The second multi-modal large language model has a larger number of model parameters than the first multi-modal large language model as a child model, is configured to review and finally determine the preliminary compliance conclusion, to generate a final compliance conclusion, and replace the preliminary compliance conclusion with the final compliance conclusion to generate the pre-audit report.

[0023] In a possible design, the pre-trained click rate prediction model is a feature interaction network model based on deep learning, and the click rate prediction model determines the prediction of the click rate by the following formula:

[0024]

[0025] wherein, is an S-shaped activation function, is a vector splicing operation, is a first feature embedding vector corresponding to a dense numerical feature in the delivery context information, is a second feature embedding vector matrix corresponding to a sparse categorical feature in the delivery context information and the intrinsic feature of the material, is a multi-layer perception network for processing the first feature embedding vector, is a cross-interactive network for modeling high-order feature interaction of the second feature embedding vector matrix, and are an output layer weight and a bias of the click rate prediction model, respectively.

[0026] In a possible design, the sparse categorical feature includes one or more of a commodity category ID, a target ad position ID, and an advertiser account ID; and the dense numerical feature includes one or more of a video length, a call-to-action appearance timestamp, and an advertiser historical click rate.

[0027] In a possible design, after generating the pre-audit report, the method further includes:

[0028] Based on the compliance status of the pre-audit report, performing a corresponding operation on the advertising material in an advertising delivery system of the e-commerce platform;

[0029] If the compliance status is passed, an accelerated review label is added to the advertising material in the advertising delivery system; if the compliance status is rejected, the violation position in the advertising material corresponding to the optimization suggestion of the pre-audit report is highlighted in the material editing interface of the advertising delivery system.

[0030] In one possible design, the method further includes performing a series of controllable visual disturbance tests on the advertising material to evaluate the robustness of the first multi-modal large language model and the click rate prediction model under brightness changes, blurring, or occlusion conditions.

[0031] In the second aspect, the embodiments of the present application propose an advertising material pre-audit system based on an e-commerce delivery scenario, comprising:

[0032] A data packet generation module is configured to obtain advertising material to be pre-audited and delivery context information associated with the advertising material, and perform multi-modal analysis on the advertising material to generate a pre-audit data packet, wherein the pre-audit data packet contains structured visual features, text features, and auditory features extracted from the advertising material;

[0033] A compliance detection module is configured to retrieve relevant advertising specifications matching the delivery context information based on a pre-constructed rule knowledge base dynamically corresponding to e-commerce platform advertising specifications, and perform compliance detection on the pre-audit data packet using a first multi-modal large language model, wherein the compliance detection is based on a call prompt containing the relevant advertising specifications as a context to generate a preliminary compliance conclusion;

[0034] An adaptability evaluation module is configured to extract intrinsic features of the material from the pre-audit data packet, wherein the intrinsic features of the material are at least part of the structured visual features, text features, and auditory features, and perform delivery adaptability evaluation on the advertising material based on the intrinsic features of the material and the delivery context information using a pre-trained click rate prediction model to generate a predicted adaptability score;

[0035] In addition, a report generation module is configured to integrate the preliminary compliance conclusion and the adaptability score to generate a pre-audit report containing the compliance status, the adaptability score, and optimization suggestions.

[0036] Advantages:

[0037] This invention provides a pre-screening method for advertising materials in e-commerce advertising scenarios. By automating the pre-screening process before ad delivery, it overcomes the limitations of existing "post-delivery review" models, which suffer from high risk, low efficiency, and susceptibility to errors in manual pre-screening. Utilizing a multimodal large language model and injecting dynamically retrieved platform advertising specifications as context into the call prompts, it achieves accurate, rapid, and strategy-aligned detection of advertising material compliance, effectively addressing frequent changes in platform rules. Furthermore, this invention innovatively introduces a pre-trained click-through rate prediction model to proactively assess the suitability of advertising materials for ad delivery, solving the pain point of existing technologies that only focus on compliance while neglecting performance. By integrating compliance conclusions with suitability scores, it provides advertisers with comprehensive decision-making support, significantly reducing the risk of materials being rejected due to violations and avoiding the waste of advertising budget on inefficient materials, thereby improving the overall efficiency and return on investment of ad delivery. Attached Figure Description

[0038] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0039] Figure 1 This is a flowchart of an advertising material pre-review method based on e-commerce advertising scenarios proposed in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of an advertising material pre-review system based on e-commerce advertising scenarios proposed in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0042] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the invention, but does not constitute a limitation on the invention.

[0043] The first aspect, reference Figure 1This invention provides a method for pre-screening advertising materials in an e-commerce advertising scenario, which can be executed on a computer device. The computer device can be a server, a personal computer (PC), or an embedded advertising material pre-screening system with sufficient computing power. This device typically includes, but is not limited to, a processor, memory, communication interface, and input / output devices. The memory stores computer program instructions, and the processor executes these instructions to implement the advertising material pre-screening method for e-commerce advertising scenarios of this invention. This method aims to conduct a comprehensive compliance and advertising suitability assessment of advertising materials before they are formally submitted to the e-commerce platform's review process, thereby mitigating risks, improving efficiency, and optimizing advertising effectiveness. The specific process of this method includes the following steps:

[0044] S101: Obtain the advertising creative to be pre-approved and the delivery context information associated with the advertising creative, and perform multimodal parsing on the advertising creative to generate a pre-approval data package.

[0045] This step is the data input and structured processing stage of the entire pre-screening process. Its purpose is to transform the diverse original advertising materials and their related information into a standardized data structure that can be uniformly processed by the computer model in subsequent steps.

[0046] S1011. This step first obtains the advertising materials to be pre-approved through a pre-defined interface. This interface can be deeply integrated with the merchant backend system of the e-commerce platform, allowing advertisers to directly submit images, videos, and other advertising materials to be processed through operations such as "one-click pre-approval" in the material library, eliminating the cumbersome process of repeated uploads. At the same time, the system also obtains the placement context information associated with the material, which is crucial for subsequent compliance judgment and suitability assessment. Placement context information may include: merchant ID, product ID, target advertising account information, product category (such as "beauty and personal care", "baby products"), pre-defined ad placement (such as "homepage recommendation stream", "live stream traffic placement"), target audience profile tags, advertising budget range, and the advertiser's historical performance data, etc.

[0047] S1012. After acquiring the original creative materials and contextual information, the system performs deep multimodal analysis on the advertising creative materials to extract the structured features contained therein. This analysis process calls different processing engines based on the modality of the creative materials (visual, auditory, text):

[0048] 1) Auditory Feature Extraction: For video advertising materials, the system utilizes an Automatic Speech Recognition (ASR) engine. ASR technology, especially end-to-end models based on deep learning, can accurately transcribe spoken audio (such as the anchor's script and background narration) into text, i.e., "spoken script text." This step is crucial for detecting verbal promises, illegal promotional language, or prohibited words, compensating for the shortcomings of relying solely on visual and textual analysis.

[0049] 2) Visual Feature Extraction: The system utilizes a computer vision (CV) engine to analyze images or video frames and extract key structured visual features. In a possible design, this step includes:

[0050] Brand logo detection identifies the presence of a brand's logo in an image and records its position (e.g., using a "nine-grid" positioning method to categorize it into areas such as "top left," "center," and "bottom right"). Optical Character Recognition (OCR) extracts all text information contained in the image, including artistic fonts, watermarks, and text on product packaging, and identifies the position and coverage area of ​​this text in the image. Product subject recognition segments and identifies the core product in the image, assessing its clarity, completeness, and visibility to ensure it meets the basic requirements for product display on e-commerce platforms. Human presence detection detects the presence of humans in the image and further identifies faces, providing foundational data for subsequent portrait rights and model compliance reviews. Dominant color analysis analyzes the main color composition of the image, extracts the top-3 dominant colors and their proportions, and uses this to determine whether it conforms to brand visual identity (VI) specifications.

[0051] 3) Text feature extraction: The system directly parses the text content submitted along with the advertising materials, including the advertising title, advertising copy, and call-to-action (CTA) statements.

[0052] S1013. Finally, the system integrates the spoken script text, structured visual features (logo and its position, OCR text, product information, etc.) extracted in the above steps, as well as the original text features, with the delivery context information obtained in S1011, and encapsulates them into a unified, structured pre-screening data package. This data package is typically stored in formats such as JSON (JavaScript Object Notation) to provide standardized input for all subsequent analysis steps.

[0053] S102: Based on a pre-built rule knowledge base that dynamically corresponds to the advertising specifications of e-commerce platforms, retrieve relevant advertising specifications that match the context information of the campaign.

[0054] The core of this step is to build a knowledge base that can accurately reflect and dynamically adapt to the complex advertising rule system of e-commerce platforms, and retrieve the most relevant rule subsets based on specific pre-screening tasks.

[0055] S1021. Construction of the Rule Knowledge Base: This method pre-constructs a comprehensive rule knowledge base. This knowledge base collects, organizes, and structures all publicly released advertising review guidelines from e-commerce platforms. In one possible design, the knowledge base is constructed using a hierarchical classification system. This system organizes the complex advertising guidelines according to a tree structure of "Domain → Topic → Subtype → Behavior". For example, a rule about "prohibiting absolute terms in cosmetic advertisements" might be categorized as "Beauty and Skincare Domain → Advertising Topic → Prohibited Terms Subtype → Absolute Terms Behavior". This hierarchical structure not only clarifies the rule system but, more importantly, supports subsequent models in performing coarse-to-fine path-based risk identification, improving the logic and accuracy of detection. The knowledge base is designed to be dynamically updated. When the platform releases new rules or modifies old rules, operations personnel can easily update the corresponding rule texts in the database to ensure that the pre-review standards are always consistent with the platform's official standards.

[0056] S1022. Retrieval of Relevant Advertising Guidelines: During each pre-review task, the system first analyzes the contextual information in the pre-review data package. For example, if the product category is detected as "medical device," the system will retrieve, in addition to the platform's general advertising guidelines, all specific rules related to the "medical device" field from the rule knowledge base. This creates a highly relevant "temporary rule set" for the current material to be reviewed, avoiding wasting computational resources on irrelevant rules and improving the targeting of the detection.

[0057] S103: Utilize the first multimodal large language model to perform compliance checks on the pre-review data packet. The compliance checks are based on call prompts that include relevant advertising specifications as context, in order to generate preliminary compliance conclusions.

[0058] This step is the core of compliance judgment. Its innovation lies in the adoption of the "Policy-Aligned" reasoning paradigm, which allows the judgment of large language models to be based directly on the latest rule text, rather than relying solely on patterns indirectly learned from labeled data.

[0059] S1031. Construction and Injection of Prompt: The system programmatically constructs a structured prompt from the relevant advertising specification text retrieved in S102. This prompt not only includes the role setting of the model (e.g., "You are an experienced e-commerce advertising review expert"), but also explicitly "injects" each advertising specification that needs to be checked as contextual information into the prompt. For example, the prompt may contain an instruction like: "Please determine whether the text content of the input material is compliant according to the following rule: [Rule 1: The advertisement shall not contain terms such as 'national level', 'highest level', 'best', etc."]

[0060] S1032. Compliance Detection for Execution Strategy Alignment: The system submits the constructed call prompts, along with the multimodal features (such as OCR text, spoken script text, and advertising copy) from the pre-review data package, to the first Multimodal Large Language Model (MLLM). Upon receiving the input, this MLLM (e.g., a model with a moderate number of parameters to balance cost and efficiency) directly performs logical reasoning and matching judgments on the content of the pre-review data package based on the rule text explicitly given in the prompts. For example, it checks whether any prohibited words listed in the rules appear in all text features. This "instant rule provision" approach ensures that every judgment of the model is based on evidence, and its decision logic is directly aligned with the platform's current policies.

[0061] S1033. Generate preliminary compliance conclusion. After the first MLLM completes the reasoning, it will output a preliminary compliance conclusion in a preset format. This conclusion usually includes an overall compliance judgment (e.g., "compliant" or "non-compliant"), as well as (in the case of non-compliance) the specific violations, the content of the violations, and the rule entries violated.

[0062] S1034. In one possible design, after the above compliance checks, this method may also include a fact-checking step based on external evidence.

[0063] S10341. The system determines whether the pre-screening data packet contains a verification statement that requires external information verification, such as "recommended by experts" or "number one in sales across the entire network".

[0064] S10342. If such a statement is included, a fact-checking process will be triggered. This process mimics the working mode of a human fact checker, dynamically planning and executing a series of verification actions. For example, for an "expert" image, a reverse image search is performed to confirm whether the image is stolen or outdated material; for a text statement that "has supporting evidence," a web search is performed to find supporting evidence published by authoritative journals or official institutions.

[0065] S10343. The system will summarize and analyze the retrieved external validation evidence and update the preliminary compliance conclusion based on this evidence. For example, if no reliable evidence supporting "clinical trial proof" is found, the item will be marked as "suspected false advertising," and the compliance conclusion will be updated.

[0066] S1035. In one possible design, to balance review costs and accuracy, this method may adopt a “parent-child model” collaborative adjudication architecture.

[0067] S10351. In this architecture, the first multimodal large language model used in S103 is regarded as a "sub-model". It is usually a lightweight model with low inference cost and fast speed, responsible for handling the vast majority of routine pre-screening tasks.

[0068] S10352. The system will evaluate the confidence level of the sub-model's preliminary compliance conclusions. If the confidence level is lower than a preset threshold (for example, for an ambiguous case, the confidence level of the sub-model's judgment may only be 60%), it indicates that the case is relatively complex or has a high risk.

[0069] S10353. At this point, the preliminary data package and initial conclusions of the case will be automatically submitted to a second multimodal large language model, which will serve as the "parent model". This parent model is typically a model with a larger number of parameters, stronger reasoning capabilities, but also higher costs.

[0070] S10354. The parent model, leveraging its superior comprehensive analytical capabilities, reviews and makes a final ruling on the conclusions of the sub-models, generating a final compliance conclusion. This final compliance conclusion then replaces the original preliminary compliance conclusion for the generation of subsequent pre-audit reports. In this way, the system only utilizes the high-cost parent model when necessary, thereby significantly reducing overall operating costs while ensuring accuracy in high-risk cases.

[0071] S104: Extract the intrinsic features of the creative materials from the pre-approval data package, and based on the intrinsic features of the creative materials and the context information of the delivery, use a pre-trained click-through rate prediction model to evaluate the delivery suitability of the advertising creative materials in order to generate a predicted suitability score.

[0072] This step aims to go beyond traditional compliance reviews, quantifying and predicting the potential of ad creatives from a market performance perspective, and providing advertisers with more commercially valuable insights.

[0073] S1041. Extraction of Intrinsic Features of Creative Materials: From the pre-screening data package generated in S101, the system further extracts intrinsic features of creative materials that are strongly correlated with the ad's CTR. These features are inherent attributes of the creative materials themselves and are at least a part of structured visual features, textual features, and auditory features. In a possible design, these features include: video duration, color saturation, editing rhythm, timestamps of CTA elements, copy length, and the emotional polarity of the spoken word, etc.

[0074] S1042. Application of the Click-Through Rate (CTR) Prediction Model: This method employs a pre-trained CTR prediction model. This model is built upon deep learning and trained on massive amounts of historical advertising data (including creative features, contextual features, and corresponding actual CTRs). During pre-screening, the system uses the extracted intrinsic features of the creative materials, along with the contextual information from the pre-screening data package (such as product category, ad placement, target audience, etc.), as input to feed the CTR prediction model.

[0075] S1043. In one possible design, the pre-trained click-through rate (CTR) prediction model is based on a deep learning feature interaction network model. The core advantage of this type of model lies in its ability to automatically learn and capture the interaction relationships between high-order, non-linear features, which is crucial for CTR prediction tasks. A specific model structure can be determined by the following formula:

[0076]

[0077] S1044. Detailed Explanation and Optimization of Formula Parameters:

[0078] : Represents the predicted click-through rate output by the model, which is a probability value between 0 and 1. : This is the sigmoid function. Its function is to map any real number to the interval (0,1), so that its output conforms to the definition of probability, which facilitates subsequent interpretation and application. and : These represent the embeddings of dense features and sparse features, respectively. Sparse categorical features are high-dimensional features with discrete values, such as product category ID, target ad slot ID, and advertiser account ID. These features are typically first encoded using one-hot encoding, and then mapped to a low-dimensional dense vector matrix through an embedding layer. Dense numerical features are continuous numerical features, such as video duration, timestamps of calls to action, and advertisers' historical average click-through rates. These features are usually normalized and can be used directly or mapped to vectors through an embedding layer. . This represents a Multi-Layer Perceptron (MLP) network. It consists of multiple fully connected layers and is specifically designed to process and extract nonlinear information from dense features. : Represents a Cross Interaction Network (CIN). This is one of the core components of the model, used to explicitly and layer-by-layer learn high-order combinatorial relationships between sparse features. Compared to traditional inner products or Hadamard products, CINs can perform feature interactions at the vector level, preserving more original information and thus more effectively capturing complex feature interaction patterns. : Represents a vector concatenation operation. It concatenates the dense feature representation after MLP processing with the sparse feature representation after CIN processing in a dimensional manner to form a wide vector containing various types of information and interactive information. and : These represent the weight matrix and bias vector of the final output layer, respectively. They perform a linear transformation on the concatenated wide vector, and finally obtain the predicted click-through rate through the Sigmoid function.

[0079] The optimization of this model lies in its parallel MLP and CIN structures, which process features of different properties separately and then fuse them. This ensures effective modeling of different features and strengthens the learning of high-order feature interactions that are crucial for CTR prediction through the CIN network. Compared with architectures that only use MLP or shallow models, it has stronger expressive power and higher prediction accuracy.

[0080] S1045, Generate prediction fit score: Model output The value is a raw probability. To make it more intuitive for users, the system converts it into a "fitness score" on a percentage scale. This is typically done by predicting... This is achieved by comparing and normalizing the CTR with the historical average CTR of the product category and ad placement, or with a benchmark CTR of top-performing creatives. For example, a predicted CTR significantly higher than the historical average might be translated into a suitability score of 90 or higher.

[0081] S105: Integrate preliminary compliance conclusions and suitability scores to generate a pre-audit report that includes compliance status, suitability score, and optimization suggestions.

[0082] This step is the final output of the entire process, designed to present all analysis results to the user in a clear and actionable format.

[0083] S1051. The system integrates the compliance conclusion generated in S103 (which may have been optimized by S1034 and S1035) with the adaptability score generated in S104.

[0084] S1052. Based on the integrated information, the system automatically generates a structured pre-review report. This report includes at least the following three core parts: Compliance Status, clearly marked as "Pass" or "Rejected." If "Rejected," it will list all violations, violation content fragments (such as timestamps of violating text or images), the specific rule clauses violated, and the decision reasons from MLLM. Adaptability Score, displaying a quantified predicted score for campaign performance, and can include comparison charts with industry / historical averages, allowing advertisers to clearly understand the potential of their creative materials. Optimization Suggestions, providing specific and actionable modification suggestions. For compliance issues, the suggestions are mandatory, such as "Please remove the phrase 'Number One Brand' from the copy"; for adaptability issues, the suggestions are instructive, such as "It is recommended to move the call-to-action button to the first 5 seconds of the video to increase click-through rates."

[0085] Through the coordinated execution of the above steps, the method disclosed in this embodiment constructs a closed-loop, intelligent pre-review process for advertising materials, which not only solves compliance risks before delivery but also provides unprecedented insights into delivery performance, offering powerful decision support for e-commerce advertisers.

[0086] For the second aspect, please refer to... Figure 2 This invention also proposes an advertising material pre-screening system based on e-commerce advertising scenarios. This system is adapted to the pre-screening method described in the foregoing embodiments and can achieve automated and intelligent comprehensive evaluation of advertising materials before advertising campaigns. The system can be deployed on a cloud server or a local server cluster, forming an integrated pre-screening framework through modular functional design. The system includes a data package generation module, a compliance detection module, a compatibility assessment module, and a report generation module, wherein each module works collaboratively, as detailed below.

[0087] The data package generation module is used to acquire advertising creatives to be pre-approved and the associated placement context information, and to perform multimodal parsing on the advertising creatives to generate a pre-approval data package. Specifically, this module connects to the merchant backend system of the e-commerce platform through an Application Programming Interface (API), receives image or video creatives submitted by advertisers, and collects associated placement context information, such as product ID, target ad position, and category. Upon receiving the input, the module integrates multiple parsing engines: for video creatives, its built-in ASR engine transcribes the spoken audio into spoken text; its CV engine is responsible for extracting structured visual features from images or video frames, including but not limited to text extracted through OCR, logos and their positions in the image, the main product area, and information about the presence of people. This module integrates all the parsed auditory, visual, and textual features with the placement context information to generate a standardized, machine-readable JSON format pre-approval data package, which serves as the unified input for all subsequent modules.

[0088] The compliance detection module is used to perform compliance checks on pre-approved data packets based on a pre-built rule knowledge base that dynamically corresponds to the e-commerce platform's advertising specifications. Specifically, the core of this module is a rule knowledge base that is updated synchronously with the e-commerce platform's advertising specifications. In one possible design, this knowledge base is built using a hierarchical classification system, organizing hundreds of complex advertising specifications according to a "domain → topic → subtype → behavior" structure to support path-based risk identification. During detection, the module first retrieves a subset of highly relevant advertising specifications from the knowledge base based on the placement context information (such as product category) in the pre-approved data packet. Then, it injects these specification texts as context into a call hint built for the first MLLM (Multi-Level Model). Upon receiving the hint containing the rule context and the pre-approved data packet, the model directly performs logical reasoning based on the explicit rules to determine whether the content violates regulations, thereby generating a preliminary compliance conclusion.

[0089] To improve the accuracy and robustness of the detection, this module can also integrate two preferred sub-functions. First, for statements to be verified (such as "expert recommendations") included in the materials, this module can invoke a fact-checking process, retrieving verification evidence from external data sources through reverse image searches, web searches, and other verification actions, and updating the compliance conclusion based on the evidence. Second, this module can adopt a "mother-child model" collaborative adjudication architecture. Under this architecture, the first MLLM acts as a "child model" handling routine tasks. When its confidence level in judging a case falls below a preset threshold, the case will be automatically reported to a "mother model" with a larger number of parameters and stronger reasoning capabilities for review and final adjudication. This design significantly optimizes the system's operating costs while ensuring the accuracy of high-risk case judgments.

[0090] The suitability assessment module extracts intrinsic features of creative materials from the pre-screening data package and, combined with the delivery context information, evaluates the suitability of ad creatives for delivery. Specifically, this module aims to quantitatively predict the potential performance of ad creatives in specific delivery scenarios, especially CTR. It first extracts intrinsic features of creative materials strongly correlated with CTR from the pre-screening data package, such as video duration, color saturation, and the timing of CTA appearance. Then, these intrinsic features, along with delivery context information (such as ad placement, product category, and advertiser historical performance), are fed into a pre-trained click-through rate prediction model. In one possible design, this model is based on a deep learning-based feature interaction network model, which can effectively capture high-order, non-linear feature interaction relationships, thereby outputting a predicted click-through rate value. Finally, the module compares the predicted click-through rate with historical data benchmarks and normalizes it into a "suitability score" of 0-100, which is then presented intuitively to the user.

[0091] The report generation module integrates preliminary compliance conclusions and suitability scores to generate a pre-review report. Specifically, this module is the system's output, responsible for summarizing all the aforementioned analysis results into a user-friendly and actionable report. The report has a clear structure, containing at least three parts: a clear compliance status (passed / rejected), along with specific violations and the violated rules; a quantified suitability score, providing a comparison with industry averages; and priority-based optimization suggestions, divided into "must-modify items" (for compliance issues) and "suggested optimization items" (for improving campaign performance). Furthermore, this module has the ability to collaborate with e-commerce platform workflows. For example, if the pre-review result is "passed," the module can send an instruction to the advertising system to add an "accelerated review" tag to the creative; if the result is "rejected," the module can automatically highlight the violation when the advertiser enters the creative editing interface, greatly improving modification efficiency.

[0092] The third aspect of this application provides a computer device for pre-screening advertising materials in an e-commerce advertising scenario. Please refer to [link / reference]. Figure 3 The device includes a memory and a processor connected in series and communicate with each other. The memory stores computer programs, and the processor reads the computer programs and executes the advertising material pre-screening method based on e-commerce placement scenarios, as described in the first aspect of this embodiment. Specifically, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or last-in-first-out (FILO) memory, etc.; the processor may be, but is not limited to, microprocessors of the STM32F105 series, ARM (Advanced RISC Machines), x86 architecture processors, or processors with integrated NPUs (neural-network processing units). The working process, working details, and technical effects of the device provided in the second aspect of this embodiment can be found in the first aspect of this embodiment, and will not be repeated here.

[0093] This fourth aspect of the embodiment provides a computer-readable storage medium storing instructions containing the advertising material pre-approval method for an e-commerce advertising scenario according to the first aspect of the embodiment. Specifically, the computer-readable storage medium stores instructions that, when executed on a computer, perform the advertising material pre-approval method for an e-commerce advertising scenario as described in the first aspect. The computer-readable storage medium refers to a data storage medium, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device systems. The working process, details, and technical effects of the computer-readable storage medium provided in this fourth aspect of the embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0094] The fifth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform an advertising material pre-screening method based on an e-commerce placement scenario, as described in the first aspect of this embodiment. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0095] The embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer or programmable device to execute the methods of various embodiments or some parts of embodiments.

[0097] Finally, it should be noted that although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for pre-auditing advertisement materials based on an e-commerce delivery scene, characterized in that, The method comprises the following steps: obtaining an advertisement material to be pre-audited and a delivery context information associated with the advertisement material, and performing multi-modal analysis on the advertisement material to generate a pre-audit data package, wherein the pre-audit data package contains structured visual features, text features and auditory features extracted from the advertisement material; based on a pre-constructed rule knowledge base dynamically corresponding to the e-commerce platform advertisement specification, retrieving relevant advertisement specifications matching the delivery context information; using a first multi-modal large language model, performing compliance detection on the pre-audit data package, wherein the compliance detection is based on a call prompt containing the relevant advertisement specifications as context, to generate a preliminary compliance conclusion; extracting internal features of the material from the pre-audit data package, wherein the internal features of the material are at least part of the structured visual features, text features and auditory features, and based on the internal features of the material and the delivery context information, using a pre-trained click-through rate prediction model to perform delivery adaptability evaluation on the advertisement material to generate a predicted adaptability score; and integrating the preliminary compliance conclusion and the adaptability score to generate a pre-audit report containing compliance status, adaptability score and optimization suggestions.

2. The method of claim 1, wherein, The step of performing multi-modal analysis on the advertisement material comprises: if the advertisement material is a video material, using automatic speech recognition to extract the voice-over audio in the video material corresponding to the voice-over text; and using computer vision to extract visual elements in the advertisement material, wherein the visual elements at least include: brand logo, text extracted by optical character recognition and product main body.

3. The method of claim 1, wherein, The rule knowledge base is constructed using a hierarchical classification system, which divides the advertisement specification into multiple levels from field, theme, sub-type to behavior to realize path-based risk identification.

4. The method according to claim 1 or 3, characterized in that, After performing compliance detection, it further comprises: determining whether the pre-audit data package contains a statement to be checked that requires external information verification; if yes, dynamically planning and performing at least one verification action to retrieve verification evidence from external data sources, and updating the preliminary compliance conclusion based on the verification evidence; the verification action includes reverse image search or web search.

5. The method of claim 1, wherein, After performing compliance detection, it further comprises: if the output confidence of the first multi-modal large language model on the preliminary compliance conclusion is lower than a preset confidence threshold, submitting the pre-audit data package and the preliminary compliance conclusion to a second multi-modal large language model as a parent model; wherein the model parameter quantity of the second multi-modal large language model is greater than that of the first multi-modal large language model as a child model, which is used to review and finally determine the preliminary compliance conclusion to generate a final compliance conclusion, and the final compliance conclusion is used to replace the preliminary compliance conclusion to generate the pre-audit report.

6. The method of claim 1, wherein, The pre-trained click rate prediction model is a deep learning based feature interaction network model, and the prediction of the click rate is determined by the following formula: wherein, is a Sigmoid activation function, is a vector concatenation operation, is a first feature embedding vector corresponding to dense numerical features in the delivery context information, is a second feature embedding vector matrix corresponding to sparse categorical features in the delivery context information and the intrinsic features of the material, is a multi-layer perceptron network for processing the first feature embedding vector, is a cross-interaction network for modeling high-order feature interactions of the second feature embedding vector matrix, and are output layer weights and bias of the click rate prediction model, respectively.

7. The method of claim 6, wherein, The sparse categorical features include one or more of a commodity category ID, a target ad position ID, and an advertiser account ID; and the dense numerical features include one or more of a video duration, a call-to-action appearance timestamp, and an advertiser historical click-through rate.

8. The method of claim 1, wherein, After the pre-audit report is generated, the method further includes: Based on the compliance status of the pre-audit report, performing a corresponding operation on the advertising material in an advertising delivery system of the e-commerce platform; If the compliance status is passed, an accelerated review mark is added to the advertising material in the advertising delivery system; if the compliance status is rejected, a violation position in the advertising material corresponding to the optimization suggestion of the pre-audit report is highlighted in a material editing interface of the advertising delivery system.

9. The method of claim 1, wherein, The method further includes performing a series of controllable visual disturbance tests on the advertising material to evaluate the robustness of the first multi-modal large language model and the click-through rate prediction model under conditions of brightness change, blurring, or occlusion. 10.A system for pre-auditing advertisement materials based on an e-commerce delivery scene, characterized in that, The method includes: a data packet generation module configured to obtain an advertising material to be pre-audited and delivery context information associated with the advertising material, and perform multi-modal analysis on the advertising material to generate a pre-audit data packet, the pre-audit data packet containing structured visual features, text features, and auditory features extracted from the advertising material; a compliance detection module configured to retrieve relevant advertising specifications matching the delivery context information based on a pre-constructed rule knowledge base dynamically corresponding to e-commerce platform advertising specifications, and perform compliance detection on the pre-audit data packet using a first multi-modal large language model, wherein the compliance detection is based on a call prompt containing the relevant advertising specifications as context to generate a preliminary compliance conclusion; an adaptability evaluation module configured to extract intrinsic features of the advertising material from the pre-audit data packet, the intrinsic features being at least a part of the structured visual features, text features, and auditory features, and perform delivery adaptability evaluation on the advertising material based on the intrinsic features of the advertising material and the delivery context information using a pre-trained click-through rate prediction model to generate a predicted adaptability score; and a report generation module configured to integrate the preliminary compliance conclusion and the adaptability score to generate a pre-audit report containing a compliance status, an adaptability score, and an optimization suggestion.

Images