An AI-generated advertising content violation risk identification method and system

Abstract

The application discloses an AI-generated advertisement content violation risk identification method and system, which extracts text, image and audio three types of modal semantic anchor points to generate structured material representation, performs directional cross-modal joint feature fusion with semantic anchor points as constraints to avoid irrelevant content dilution of violation clues; the anchor point guided fusion feature sequence is input into a violation risk classifier to identify independently by category to obtain preliminary risk identification results with anchor point positioning; a risk attribution report is generated through business scene weighted scoring and cross-modal attribution decomposition to quantitatively decompose the comprehensive risk score to each target semantic anchor point; based on cross-modal consistency, the confidence of the anchor point supported by a single modal isolated evidence is reduced and the score is recalculated, and the modified comprehensive risk score, risk level and updated risk attribution report are output. The application significantly improves the traceability and audit credibility of the AI-generated advertisement violation identification conclusion.

Description

Technical Field

[0001] This invention relates to the field of risk identification technology, specifically to a method and system for identifying violations of AI-generated advertising content. Background Technology

[0002] With the large-scale application of AI-generated technology in the advertising industry, the text, visuals, and voiceovers of advertising materials can all be generated in batches by AI models, significantly increasing the speed and quantity of advertising content production. To prevent the risks of false advertising, misleading consumption, and prohibited content that may be included in AI-generated ads, existing review solutions typically employ multimodal content understanding models. These models input the text, image frames, and audio of the advertising materials into a unified manner for global feature fusion, and then the classification head outputs a judgment on whether the entire material violates regulations. Some solutions also provide a violation confidence score for the entire material after determining a violation, allowing reviewers to sort and process it.

[0003] However, the aforementioned globally integrated review scheme has significant flaws: the model processes all materials indiscriminately, diluting scattered violation clues in the text, visuals, and voiceovers with a large amount of irrelevant content. The output review conclusion is only a general judgment on the entire material, unable to trace back to the specific character range, visual area, or audio time interval where the violation clue appeared. Reviewers, faced with a material deemed in violation, have no way of knowing what specific evidence supports the conclusion, making manual verification difficult and providing traceable judgment basis when the reviewed party appeals. This results in insufficient interpretability and verifiability of violation identification conclusions, hindering review efficiency and credibility. Summary of the Invention

[0004] This invention aims to provide a method and system for identifying violations of AI-generated advertising content, which significantly improves the traceability and credibility of the identification results of violations of AI-generated advertising.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a method for identifying the violation risks of AI-generated advertising content, comprising: It receives AI-generated multimodal advertising materials, separates them into text stream, image stream, and audio stream, extracts semantic anchors of the three modalities, and generates structured material representations; Using semantic anchors as constraints, directional cross-modal joint feature fusion is performed on the structured material representation to obtain an anchor-guided fused feature sequence; The fusion feature sequence guided by the anchor point is input into the violation risk classifier, and the violation risk is independently identified for each category to obtain the preliminary risk identification result with anchor point positioning. Based on the business scenario weights, a comprehensive risk score is calculated and a risk level is determined for the preliminary risk identification results; The comprehensive risk score is subjected to cross-modal attribution decomposition to generate a risk attribution report; Based on cross-modal consistency, the identification results are subjected to secondary verification and confidence correction, and the corrected comprehensive risk score, corrected risk level and updated risk attribution report are output.

[0006] Preferably, the extraction of three types of modal semantic anchors includes: The text stream is scanned according to the text rule set configured based on the preset set of violation risk categories. The sentence fragments that hit the rules are marked as text semantic anchors, and the anchor number, associated risk category number and character-level start and end positions are recorded. Text recognition is performed on the image stream. The text on the screen is compared with the text semantic anchors according to the bidirectional character overlap rate. When the preset standard is met, the anchor number of the corresponding text semantic anchor is used. The visual symbol detector independently establishes image semantic anchors for non-text visual elements according to the preset illegal visual symbol library. Anchors with the same associated risk category number and whose display time overlaps with the display time of the subtitle corresponding to the text semantic anchor are merged. If there are multiple candidates, the one with the longest overlap time is merged. The audio stream is converted from speech to text to obtain narration text. The anchor point number of the text semantic anchor point is used according to the bidirectional word overlap rate. The audio feature discriminator independently establishes audio semantic anchor points for suspicious acoustic segments based on spectral features and processes them according to the same merging rules. The merged three types of modal semantic anchors are mapped to a unified time axis, and the structured material representation is output.

[0007] Preferably, performing directed cross-modal joint feature fusion on the structured material representation includes: The three contents in the structured material representation are vectorized respectively and arranged into three initial feature sequences; Generate a three-way mask vector for each semantic anchor point one by one, and simultaneously select one from the masks of multiple semantic anchor points that fall into the same feature unit. The similarity between feature units is calculated and divided by a preset scaling factor before an exponential transformation is performed. Only feature units that are simultaneously activated by the mask of the same semantic anchor point are allowed to interact. After row normalization, the cross-modal alignment weight with anchor point index is obtained. When a certain modality combination has no covered unit on the semantic anchor point, all cross-modal alignment weights under the modality combination are set to zero. Each of the three modal pooling vectors is concatenated with the enhancement terms in the three directions one by one according to the anchor point. The resulting fusion vectors of each semantic anchor point are obtained by linear mapping compression and arranged in the order of the anchor point numbers to form the fusion feature sequence guided by the anchor points.

[0008] Preferably, the step of independently identifying violation risks by category includes: The fusion vector of each semantic anchor in the fusion feature sequence guided by the anchor is input into the violation risk classifier to obtain the original discriminant value of each semantic anchor in each risk category; Each original discriminant value is normalized independently to obtain the anchor-level confidence score, and the anchor-level confidence scores of each category are not diluted. For each risk category, all semantic anchors with anchor-level confidence scores not lower than the preset anchor-level threshold are included in the support anchor set. The highest anchor-level confidence score under the category is used as the material-level confidence score. Categories with material-level confidence scores higher than the preset category judgment threshold are included in the initial hit risk set. The category judgment threshold is set higher than the anchor-level threshold. Integrate the three types of spatiotemporal positions of each support anchor point, and output the preliminary risk identification results with anchor point positioning.

[0009] Preferably, the calculation of the comprehensive risk score and the determination of the risk level include: Based on the risk category number of each hit category in the preliminary hit risk set, query the preset business scenario and risk weight mapping table to obtain the business weight coefficient of each hit category in the current deployment scenario; The normalized business weight is obtained by dividing the business weight coefficient of each hit category by the sum of the business weight coefficients of all hit categories. The normalized business weights are multiplied by the corresponding category's material-level confidence scores, and then summed over all hit categories to obtain the comprehensive risk score. The risk level is determined by comparing the comprehensive risk score with the score threshold defined by the quantile of the score distribution based on the historical material review records. If the confidence level of any material that hits a category is not lower than the preset severity threshold and belongs to the preset strong regulatory hazard level, it is directly defined as high risk. Materials with a low risk level are terminated in advance, and the comprehensive risk score and risk level are used as the final output.

[0010] Preferably, the cross-modal attribution decomposition includes: Take the union of all support anchor points of the hit categories and remove duplicates to form the target semantic anchor point set; Using the arithmetic mean of the fusion vectors of all semantic anchors as the baseline vector, the fusion vector of each target semantic anchor is replaced one by one with the baseline vector. The entire process of independently identifying violation risks by category, calculating the comprehensive risk score and defining the risk level is re-executed with the fusion feature sequence guided by the replaced anchors, so as to obtain the post-perturbation comprehensive risk score of each target semantic anchor. The attribution contribution ratio of each target semantic anchor is obtained by dividing the absolute value of the difference between the comprehensive risk score and each post-disturbance comprehensive risk score by the sum of the differences of all target semantic anchors. When all differences are zero, the attribution contribution ratio of each target semantic anchor is recorded as the reciprocal of the number of target semantic anchors, and evidence redundancy is marked.

[0011] Preferably, the hierarchical generation of the risk attribution report includes: The first layer records the unique identifier, risk level, and comprehensive risk score of the material. The second layer lists the attribution contribution percentage, associated risk category number, and spatiotemporal location of each target semantic anchor point according to the anchor point number. The third layer identifies target semantic anchors whose attribution contribution exceeds a preset display threshold, and uses natural language to summarize the modalities and content segments where violation clues appear, as well as the risk category conclusions they support.

[0012] Preferably, the secondary verification includes: Select target semantic anchors whose attribution contribution ratio exceeds the preset verification trigger threshold as objects to be verified. For each object to be verified, the arithmetic mean of the original similarities of all feature pairs within the coverage of each modality combination is taken as the cross-verification strength of the modality combination. The original similarities are not normalized and aggregated and are independent of the number of coverage units of any modality. The consistency variance is calculated based on the interaction verification strength of the three modal combinations. When the consistency variance is not lower than a preset consistency threshold, the corresponding object to be verified is marked as a cross-modal inconsistency anchor point.

[0013] Preferably, the confidence correction includes: Using the consistency threshold and the preset upper limit of variance as interpolation boundaries, the attenuation factor is obtained by linear interpolation according to the consistency variance. The larger the variance, the smaller the attenuation factor. The anchor-level confidence scores of all risk categories of the cross-modal inconsistency anchor itself are multiplied by the attenuation factor, while the anchor-level confidence scores of the remaining semantic anchors remain unchanged. The two-level judgment and scoring classification are re-executed with the revised anchor-level confidence level to obtain the revised comprehensive risk score and the revised risk level. The risk attribution report is updated synchronously with the first-level level and score record based on the correction results. The cross-modal inconsistency anchor points that trigger the correction are annotated with the confidence level, score, consistency variance and weak mode combination before and after the correction, and the updated risk attribution report is output.

[0014] On the other hand, this invention proposes a system for identifying violations of AI-generated advertising content, comprising: The receiving unit is used to receive AI-generated multimodal advertising materials, separate them into text stream, image stream and audio stream, extract semantic anchors of the three modalities, and generate structured material representations; The fusion unit is used to perform directional cross-modal joint feature fusion on the structured material representation with semantic anchors as constraints, so as to obtain an anchor-guided fusion feature sequence; The classification unit is used to input the fusion feature sequence guided by the anchor point into the violation risk classifier, independently identify the violation risk by category, and obtain the preliminary risk identification result with anchor point positioning; The scoring unit is used to calculate a comprehensive risk score and determine the risk level based on the business scenario weights of the preliminary risk identification results; The attribution unit is used to perform cross-modal attribution decomposition on the comprehensive risk score and generate a risk attribution report; The verification unit is used to perform secondary verification and confidence correction on the identification results based on cross-modal consistency, and output the corrected comprehensive risk score, the corrected risk level, and the updated risk attribution report.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention extracts semantic anchors from text, images, and audio to generate structured material representations, enabling violation clues to be located to specific content segments within a specific modality before feature fusion. It performs targeted cross-modal joint feature fusion constrained by semantic anchors, ensuring cross-modal interaction only occurs between feature units covered by the anchors, preventing irrelevant content from diluting violation clues. The anchor-guided fused feature sequence is input into a violation risk classifier for independent category-by-category identification, yielding preliminary risk identification results with anchor-based localization. Each risk conclusion naturally carries the number and location of the supporting anchor. Information is then processed through business scenario-weighted scoring and cross-modal attribution decomposition to generate a risk attribution report. The comprehensive risk score is quantitatively decomposed into each target semantic anchor point, clarifying the contribution ratio of each anchor point to the score and its associated risk category and spatiotemporal location in the three modalities. Finally, based on cross-modal consistency, a second verification and confidence correction are performed. The confidence of anchor points supported by only a single modality is reduced and the score is recalculated. The corrected comprehensive risk score, corrected risk level, and updated risk attribution report are output, significantly improving the traceability and credibility of AI-generated advertising violation identification conclusions. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the method for identifying violations of advertising content generated by AI according to the present invention; Figure 2 This is a block diagram of the AI-generated advertising content violation risk identification system of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly.

[0019] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0020] Please see Figure 1 This invention provides a method for identifying the violation risks of AI-generated advertising content, including: The system receives AI-generated multimodal advertising creatives, separates them into text, image, and audio streams, extracts semantic anchors for the three modalities, and generates a structured representation of the creatives. Specifically, this includes: The text stream is scanned based on the text rule set configured according to the preset set of violation risk categories. The sentence fragments that hit the rules are marked as text semantic anchors, and the anchor number, associated risk category number and character-level start and end positions are recorded. Text recognition is performed on the image stream. The text on the screen is compared with the semantic anchor points of the text according to the bidirectional character overlap rate. When the preset standard is met, the anchor point number of the corresponding text semantic anchor point is used. The visual symbol detector independently establishes image semantic anchor points for non-text visual elements according to the preset illegal visual symbol library. For anchor points with the same associated risk category number and whose display time overlaps with the display time of the subtitle corresponding to the text semantic anchor point, they are merged. If there are multiple candidates, the one with the longest overlap time is merged. The audio stream is converted from speech to text to obtain narration text. The anchor point numbering of the text semantic anchor points is adopted according to the bidirectional word overlap rate. The audio feature discriminator independently establishes audio semantic anchor points for suspicious acoustic segments based on spectral features and processes them according to the same merging rules. The merged three types of modal semantic anchor points are mapped to a unified time axis to output a structured material representation.

[0021] This helps to locate potential violation clues to specific content segments in each modality before feature fusion, and through the cross-modal anchor numbering and time overlap merging mechanism, it enables the association and integration of relevant clues scattered in different modalities, reducing information redundancy caused by modal fragmentation or repetitive representation in subsequent processing, and providing a clearer and more structured material representation basis for targeted fusion and risk attribution.

[0022] Using semantic anchors as constraints, directional cross-modal joint feature fusion is performed on the structured material representation to obtain an anchor-guided fused feature sequence; specifically including: The three contents in the structured material representation are vectorized separately and arranged into three initial feature sequences; Generate a three-way mask vector for each semantic anchor point one by one, and simultaneously select one from the masks of multiple semantic anchor points that fall into the same feature unit. The similarity between feature units is calculated and divided by a preset scaling factor before an exponential transformation is performed. Only feature units that are simultaneously activated by the mask of the same semantic anchor point are allowed to interact. After row normalization, the cross-modal alignment weight with anchor point index is obtained. When a certain modal combination has no covered unit on the semantic anchor point, all cross-modal alignment weights under the modal combination are set to zero. Each of the three modal pooling vectors is concatenated with the enhancement terms in three directions one by one according to the anchor point. After linear mapping compression, the fusion vector of each semantic anchor point is obtained and arranged in the order of anchor point number to form the anchor point-guided fusion feature sequence.

[0023] This approach confines cross-modal feature interactions to the scope covered by semantic anchors, helping to avoid the dilution of violation clue signals by irrelevant content during global fusion and making the fused features more focused on risk-related modal segments. At the same time, generating fusion vectors one by one according to anchors and retaining anchor indexes ensures that the fusion results maintain a corresponding structural relationship with semantic anchors. This is beneficial for risk conclusions to naturally carry anchor location information during subsequent classification and recognition, and also provides a relatively stable analysis unit for attribution decomposition.

[0024] The anchor-guided fusion feature sequence is input into the violation risk classifier, and violation risks are independently identified for each category, resulting in preliminary risk identification results with anchor point localization; specifically including: The fusion vector of each semantic anchor in the anchor-guided fusion feature sequence is input into the violation risk classifier to obtain the original discriminant value of each semantic anchor in each risk category; each original discriminant value is normally normalized independently to obtain the anchor-level confidence, and the anchor-level confidence of each category is not diluted. For each risk category, all semantic anchors with anchor-level confidence scores not lower than the preset anchor-level threshold are included in the support anchor set. The highest anchor-level confidence score under the category is used as the material-level confidence score. Categories with material-level confidence scores higher than the preset category judgment threshold are included in the preliminary hit risk set. The category judgment threshold is set higher than the anchor-level threshold. The three-modal spatiotemporal positions of each support anchor are integrated to output the preliminary risk identification results with anchor positioning.

[0025] Independently classifying each risk category helps avoid mutual interference or dilution of confidence between different categories; using the maximum value of the anchor-level confidence level rather than the mean value as the material-level confidence level helps preserve the risk signal of a single strong anchor point from being diluted by a large number of low-risk anchor points; the dual-threshold design of the anchor-level threshold and the category determination threshold provides a relatively flexible adjustment space for the balance between sensitivity and false alarm rate; at the same time, the output results naturally integrate the three types of modal spatiotemporal locations supporting the anchor points, so that the identification conclusions and specific evidence maintain a traceable correspondence, which facilitates subsequent attribution analysis and manual review.

[0026] Based on the weighting of business scenarios, a comprehensive risk score is calculated and a risk level is assigned to the preliminary risk identification results; specifically including: Based on the risk category number of each hit category in the preliminary hit risk set, query the preset business scenario and risk weight mapping table to obtain the business weight coefficient of each hit category in the current deployment scenario; The normalized business weight is obtained by dividing the business weight coefficient of each hit category by the sum of the business weight coefficients of all hit categories. The comprehensive risk score is obtained by multiplying the normalized business weight by the corresponding category's material-level confidence score and then summing the results across all hit categories. The risk level is determined by comparing the comprehensive risk score with the scoring threshold determined by the quantile of the score distribution based on the historical material review records. If the confidence level of any material that hits a category is not lower than the preset severity threshold and belongs to the preset strong regulatory hazard level, it is directly classified as high risk. Materials with a low risk level are terminated in advance, and the comprehensive risk score and risk level are used as the final output.

[0027] Introducing business weights based on the placement scenario allows the overall score to reflect the regulatory emphasis on risk categories in different scenarios, helping to obtain differentiated risk assessments for the same material in different business environments. Normalized weight processing reduces the impact of the number of hit categories on the scoring scale, making the scoring results more stable and comparable. Defining level thresholds based on historical review distribution helps the scoring thresholds adapt dynamically to the actual data distribution. At the same time, setting a mechanism to directly access high risk for strongly regulated categories provides a relatively reliable fallback judgment for seriously non-compliant content. Early termination of low-risk materials also helps reduce unnecessary computational overhead and improve overall review efficiency.

[0028] Perform cross-modal attribution decomposition on the comprehensive risk score to generate a risk attribution report; wherein, cross-modal attribution decomposition includes: Take the union of all support anchor points of the hit categories and remove duplicates to form the target semantic anchor point set; Using the arithmetic mean of the fusion vectors of all semantic anchors as the baseline vector, the fusion vector of each target semantic anchor is replaced one by one with the baseline vector. The entire process of independently identifying violation risks by category, calculating the comprehensive risk score and defining the risk level is re-executed with the fusion feature sequence guided by the replaced anchors, so as to obtain the post-perturbation comprehensive risk score of each target semantic anchor. The attribution contribution ratio of each target semantic anchor is obtained by dividing the absolute value of the difference between the comprehensive risk score and each post-disturbance comprehensive risk score by the sum of the differences of all target semantic anchors. When all differences are zero, the attribution contribution ratio of each target semantic anchor is recorded as the reciprocal of the number of target semantic anchors, and evidence redundancy is marked.

[0029] Furthermore, the tiered generation of risk attribution reports includes: The first layer records the unique identifier, risk level, and comprehensive risk score of the material. The second layer lists the attribution contribution percentage, associated risk category number, and spatiotemporal location of each target semantic anchor point according to the anchor point number. The third layer identifies target semantic anchors whose attribution contribution exceeds a preset display threshold, and uses natural language to summarize the modalities and content segments where violation clues appear, as well as the risk category conclusions they support.

[0030] By replacing the target semantic anchors one by one and recalculating the comprehensive risk score, the contribution of each anchor to the score result is quantified and decomposed, which helps to apply the overall risk conclusion to specific modal content segments and provide traceable evidence for review and verification. Distributing all differences equally and marking evidence redundancy when all differences are zero helps to highlight the redundancy between related clues, avoiding misinterpretation of attribution results. The hierarchical report structure takes into account material-level overviews, anchor-level details, and natural language summaries, allowing users with different needs to obtain what they need from risk overviews, evidence location, to popular descriptions. To a certain extent, it reduces the reliance on the internal principles of the model, making the review conclusions easier to understand and verify.

[0031] Based on cross-modal consistency, the identification results are subjected to secondary verification and confidence correction, and the corrected comprehensive risk score, corrected risk level and updated risk attribution report are output.

[0032] The secondary verification includes: selecting target semantic anchors whose attribution contribution ratio exceeds a preset verification trigger threshold as verification objects; for each verification object, taking the arithmetic mean of the original similarities of all feature pairs within the coverage of each modality combination as the interaction verification strength of the modality combination, wherein the original similarities are not normalized and aggregated and are independent of the number of coverage units of any modality; calculating the consistency variance based on the interaction verification strength of the three modality combinations, and marking the corresponding verification object as a cross-modality inconsistent anchor when the consistency variance is not lower than a preset consistency threshold.

[0033] Specifically, confidence level correction includes: Using the consistency threshold and the preset upper limit of variance as interpolation boundaries, the attenuation factor is obtained by linear interpolation based on the consistency variance. The larger the variance, the smaller the attenuation factor. The anchor-level confidence scores of all risk categories of cross-modal inconsistent anchors are multiplied by a decay factor, while the anchor-level confidence scores of other semantic anchors remain unchanged. The two-level judgment and scoring classification are re-executed with the revised anchor-level confidence level to obtain the revised comprehensive risk score and the revised risk level. The risk attribution report is updated synchronously with the first-level level and score records based on the correction results. For cross-modal inconsistency anchors that trigger correction, the confidence level, score, consistency variance and weak mode combination before and after correction are noted, and the updated risk attribution report is output.

[0034] By performing cross-modal consistency verification on high-contribution anchor points, it is helpful to discover and correct the problem of high confidence caused by single modal feature deviation or inter-modal information conflict, making risk assessment more robust; by only performing local confidence attenuation on cross-modal inconsistent anchor points, the possibility of misjudgment can be reduced while avoiding excessive interference with the global recognition results as much as possible. In addition, recording consistency variance, weak mode combinations, and key values ​​before and after correction in the report simultaneously enhances the transparency of the verification process, makes it easier for auditors to intuitively understand the evolution logic of the conclusions, and further improves the practical reference value of the attribution report.

[0035] On the other hand, this invention proposes a system for identifying the violation risks of AI-generated advertising content, such as... Figure 2 As shown, it includes: The receiving unit is used to receive AI-generated multimodal advertising materials, separate them into text stream, image stream and audio stream, extract semantic anchors of the three modalities, and generate structured material representations; The fusion unit is used to perform directional cross-modal joint feature fusion on the structured material representation with semantic anchors as constraints, so as to obtain an anchor-guided fusion feature sequence; The classification unit is used to input the fusion feature sequence guided by the anchor point into the violation risk classifier, independently identify the violation risk by category, and obtain the preliminary risk identification result with anchor point positioning; The scoring unit is used to calculate a comprehensive risk score and determine the risk level based on the business scenario weights of the preliminary risk identification results; The attribution unit is used to perform cross-modal attribution decomposition on the comprehensive risk score and generate a risk attribution report; The verification unit is used to perform secondary verification and confidence correction on the identification results based on cross-modal consistency, and output the corrected comprehensive risk score, the corrected risk level, and the updated risk attribution report.

[0036] Furthermore, the aforementioned units, during execution, are also used to implement other steps of the aforementioned method for identifying violations of AI-generated advertising content, as follows: Step 1: Receive AI-generated multimodal advertising creatives and extract semantic anchors for violation detection, specifically including: Step 1.1: After the AI-generated content is completed, the multimodal ad creative package is submitted to the recognition side via the data interface.

[0037] The receiver parses the packet body and separates three streams of raw data: text stream, image stream, and audio stream. The text stream includes advertising copy, subtitle files, and embedded promotional text in the video. The subtitle files display the start and end times of each sentence. The image stream is a sequence of video frames arranged in time. The audio stream is a waveform signal that combines voice-over, background music, and sound effects. It also registers a unique identifier for each piece of material and a timestamp (in seconds) for data collection, used to link outputs from different stages back to the same material.

[0038] Step 1.2: Extract semantic anchors from the text stream independently. Scan the text according to the text rule set configured according to the preset set of violation risk categories. Each rule in the rule set corresponds to a risk category number. For example, under the false advertising category, absolute terms templates and efficacy promise patterns are configured; under the inducement to consume category, time-limited urgent language and rebate promise patterns are configured; under the sensitive information leakage category, character matching rules such as ID numbers and contact information are configured. The rule set can be supplemented as the category set expands.

[0039] First, the text stream is segmented into sentences using punctuation marks as boundaries. Then, each sentence is compared with each rule: template-based rules are compared word by word using a sliding window, while character-based rules are matched position by position. Matched sentence fragments are marked as semantic anchors, and the anchor number, associated risk category number, and character-level start and end positions are recorded.

[0040] Step 1.3: Extract semantic anchor points from the image stream, including two independent paths and one merging.

[0041] Method 1: Perform text recognition on the frame sequence to obtain the text and its pixel position in the image, and compare it with the content of each text anchor point in step 1.2. The comparison adopts a two-way word overlap rate, that is, calculate the percentage of overlapping words with the number of anchor point words and the number of text words in the image respectively, and take the larger value of the two. When the larger value reaches the preset standard (e.g., 0.8), the text area in the image is marked as an image anchor point and the corresponding text anchor point number is used. The two are combined into a cross-modal anchor point pair.

[0042] Method 2: For visual elements not presented in text, a pre-trained visual symbol detector is used to detect them according to a preset library of illegal visual symbols, such as misleading price tag styles, comparison illustrations that exaggerate efficacy, and unauthorized certificate images. Once detected, an image anchor point is independently established and carries the corresponding risk category number from the symbol library. Its establishment is not contingent on the existence of an anchor point on the text side.

[0043] Merging rules: If an independent image anchor obtained through method two has the same associated category number as a text anchor and its display period overlaps with the display period of the corresponding subtitle of that text anchor, then the two are merged into the same anchor number, with the union of coverage and associated categories. When an image anchor overlaps with multiple text anchors simultaneously, the one with the longest overlap duration is merged. Anchors without overlap remain independent. Each image anchor records a pixel-level bounding box, defined by four values: the horizontal and vertical coordinates of the upper left and lower right corners, in pixels.

[0044] Step 1.4: Extract audio semantic anchors from the audio stream in Step 1.1 and complete timeline unification, which also includes two independent paths and one merging.

[0045] Method 1: Perform speech-to-text conversion on the audio stream to obtain narration text with start and end times for each sentence. Compare the text anchor points with the bidirectional word overlap rate standard in step 1.3. Segments that meet the standard will retain the corresponding anchor point number.

[0046] Approach 2: For segments lacking narration but with questionable acoustic features, such as fast-paced countdown sound effects or promising intonation segments, a pre-trained audio feature discriminator detects them based on spectral features and independently anchors them, carrying the category number configured by the discriminator. The merging rules are the same as in step 1.3, based on the condition that the associated categories are the same and the time intervals overlap, and when there are multiple candidates, the one with the longest overlap time is merged.

[0047] Subsequently, all three types of anchor points are mapped onto a unified timeline measured in seconds, with the starting point of the material playback as zero. This step produces a structured representation of the material: three streams of raw data, all semantic anchor points and their respective numbers, a set of associated category numbers, and the coverage of the three modalities (text character ranges, image bounding boxes and display periods, and audio time ranges). The modal coverage of each anchor point can be one, two, or all three categories, determined by the actual results of anchor building and merging. This representation carries prior location information about violations.

[0048] Step 2: Perform targeted cross-modal joint feature fusion with semantic anchors as constraints; specifically including: Step 2.1: Vectorize the three contents in the structured material representation respectively.

[0049] Text is segmented into fragments based on anchor point boundaries and sentences; images are divided into regions based on anchor point bounding boxes and a fixed stride; audio is segmented into segments based on anchor point time intervals and a fixed frame length. Each unit is mapped to a fixed-dimensional feature vector by a pre-trained corresponding feature extractor. The dimension is a positive integer (e.g., 512 dimensions), and each component is a normalized dimensionless value. The three channels are arranged into an initial feature sequence according to their order of appearance. Each feature unit retains its attribution information: its modality, source time period, and all anchor point numbers it falls into.

[0050] Step 2.2: Generate masks one by one according to anchor points. For each anchor point, number it q, and generate three-way mask vectors. Each vector has the same length as the initial feature sequence of that mode, and the i-th bit takes a value if and only if the feature unit falls within the coverage area of ​​the corresponding mode of anchor point q. For text units, check if their segments fall within the character range of q; for image units, check if their regions intersect with the bounding box of q; for audio units, check if their time segments overlap with the time range of q, and set the remaining bits to zero.

[0051] When a feature cell falls into multiple anchor points simultaneously, one of the masks for these anchor points is selected. That is, the mask is generated independently for each anchor point, allowing the sharing of feature cells. The mask value is a dimensionless label. Let A_a(q) be the set of covering cells of anchor point q in mode a.

[0052] Step 2.3: Input the three initial feature sequences and the mask generated by the anchor point into the anchor point-aware cross-attention layer in the multimodal fusion network. This layer only allows feature units that are activated by the same anchor point mask to interact, and the strength of the interaction is measured by the alignment weight.

[0053] First, define the similarity between two feature units: Take any i-th feature vector of mode a and j-th feature vector of mode b, divide their inner product by the product of their respective L2 norms, and the resulting ratio is denoted as the similarity. This similarity is dimensionless and ranges from -1 to 1. This similarity is not subject to any normalization or aggregation; its value depends only on the two vectors themselves and is independent of the number of covering units in either mode. The alignment weight of the i-th unit of mode a to the j-th unit of mode b at anchor point q is calculated using the following formula: ; In the formula: a and b represent one of text, image, and audio respectively, and a is not equal to b; q is the anchor point number; The aforementioned similarity is dimensionless; The mask value for anchor point q in step 2.2 is zero or one, and is dimensionless; Let q be the set of covering units on mode b; The preset scaling factor is dimensionless and takes a positive number between zero and one (e.g., 0.1). The smaller the value, the more the weight is concentrated on feature pairs with high similarity. The alignment weights are dimensionless. For any modal a cell covered by anchor point q (i.e., the one with the mask selected), the sum of its weights for all cells covered by modal b is always one.

[0054] The preset scaling factor is a normalized adjustment parameter used to scale the similarity, adjusting the sharpness of the cross-modal alignment weight distribution after exponential transformation: when the preset scaling factor is small, the distribution after exponential transformation tends to be sharp, and the cross-modal alignment weights are concentrated between the cross-modal feature pairs with the highest similarity, suppressing noise introduced by weakly correlated feature pairs; when the preset scaling factor is large, the distribution tends to be flat, and the feature pairs with the second highest similarity also obtain non-negligible alignment weights. Those skilled in the art can select a specific value of the preset scaling factor between zero and one according to the requirements of the degree of focus of cross-modal evidence.

[0055] Boundary convention: When When the set is empty or the sum of the denominators is zero (anchor point q has no coverage on modality b), all alignment weights under that modality combination are set to zero. In subsequent fusion, the corresponding enhancement terms are placed in zero vectors, and the calculation is not interrupted. The weights of all anchor points and all modality combinations are summarized into a cross-modal alignment weight matrix with anchor point indices, while retaining the original similarity values ​​of each feature pair.

[0056] Step 2.4: Based on the weight matrix from Step 2.3, fuse the weights one by one according to the anchor points. For anchor point q: First, the arithmetic mean of the feature vectors of the text, image, and audio units covered by q is calculated to obtain three modal pooling vectors. When there is no coverage for each modality, a zero vector is used to occupy the position. Secondly, calculate the enhancement terms in three directions. The text-to-image enhancement term is as follows: for each text unit covered by q, calculate the weighted sum of the feature vectors of the image units covered by q with the alignment weight of the row where the unit is located, and then take the average of the weighted sums of all text units. The text-to-audio and image-to-audio directions are calculated with the same caliber. When the weight of any modality combination is set to zero according to the boundary convention, the corresponding enhancement term takes the zero vector. Third, the three modal pooling vectors and the three directional enhancement terms are concatenated in a predetermined order to obtain a concatenated vector with a dimension six times that of a single vector (e.g., 3072). This concatenated vector is then compressed to a uniform dimension (e.g., 512) by a linear mapping with the coefficients obtained from training. The result is the fused vector of the anchor point q, and the components are still dimensionless values.

[0057] After traversing all anchor points, each fusion vector is arranged into a fusion feature sequence guided by the anchor point number. Each fusion vector is bound to the anchor point number, the set of associated category numbers, and the spatiotemporal location of the three modalities, which is equivalent to packaging the evidence of a violation clue in each modality into a vector.

[0058] Step 3: Input the fused features into the violation risk classifier to independently identify violation risks category by category; specifically including: Step 3.1: Feed the fused feature sequences one by one into the pre-trained violation risk classifier.

[0059] This classifier is a multi-label classification network trained with labeled multimodal advertising samples. The output layer nodes are set according to a preset set of violation risk categories. Each category is judged independently and does not compete with each other, allowing multiple categories to be established simultaneously for the same material.

[0060] The fusion vector of each anchor point yields a real number at each node of the output layer, called the original discriminant value of that anchor point in that category. This value can be positive or negative, is dimensionless, and all anchor points and all categories' original discriminant values ​​are completely preserved without any selection.

[0061] Step 3.2: Perform logistic normalization on the original discriminant values ​​for each anchor point and each category independently: divide by one and sum the exponents of the original discriminant value (taken as the negative of the original discriminant value). The resulting ratio is called the anchor-level confidence score, which is dimensionless and ranges from zero to one. The anchor-level confidence scores for each category are independent of each other, and the sum is not required to be a constant. Enhancement of evidence in one category will not lower the confidence scores of other categories.

[0062] Step 3.3: Determine the hit rate in two stages.

[0063] The first level determines support anchors by category: all anchors in a certain category with an anchor level confidence level not lower than the preset anchor level threshold (e.g., 0.5) are included in the support anchor set of that category, with no upper limit on the number.

[0064] The second level determines the material-level conclusion: the material-level confidence score of a certain category is taken as the largest among all anchor-level confidence scores of that category; it is compared with the preset category judgment threshold, and categories that exceed the threshold are included in the preliminary hit risk set.

[0065] Category determination thresholds are configured separately for each category, based on the regulatory severity level classification and historical false alarm rate statistics. Categories with high harm and severe penalties have lower thresholds, and all category determination thresholds are higher than the anchor point level threshold, thus ensuring that every hit category has at least one supporting anchor point. The number of categories in the hit set is denoted as N, where N is a positive integer not less than 1, and is dynamically determined by the judgment results.

[0066] Example calculation: Suppose the category set contains five categories. The original discriminant values ​​and corresponding anchor-level confidence scores of four anchor points of a certain material in the categories of false advertising and misleading consumption are as follows: Anchor point 1: False advertising 0.95 corresponds to 0.72, and misleading consumption -0.2 corresponds to 0.45; Anchor point two: False advertising 0.35 corresponds to 0.59, and misleading consumption 0.55 corresponds to 0.63; Anchor point 3: False advertising (-0.08) corresponds to 0.48, and misleading consumption (-1.2) corresponds to 0.23; Anchor point four: False advertising (-2) corresponds to 0.12, and misleading consumption (0.15) corresponds to 0.54; The confidence levels for the other three categories of materials are 0.30, 0.12, and 0.20, respectively.

[0067] Let the anchor-level threshold be 0.5, the thresholds for false advertising and misleading consumption be 0.6, and the thresholds for the other categories be 0.55. Then, the supporting anchors for false advertising are anchors one and two, with a material-level confidence level of 0.72, which is higher than 0.6 and thus has a hit rate. The supporting anchors for misleading consumption are anchors two and four, with a material-level confidence level of 0.63, which is higher than 0.6 and thus has a hit rate. The other three categories are all below the threshold and are excluded. N equals 2.

[0068] Step 3.4: Based on the set of support anchors for each category, return to the structured material representation to extract the text character range, image bounding box, and audio time range of each support anchor, and integrate the positioning records that can contain multiple anchors for each hit category.

[0069] The preliminary risk identification results with anchor point positioning are as follows: unique identifier of the material, number of each hit category and material-level confidence level, complete set of supporting anchor points for each category, and spatiotemporal location of the three modalities. This result provides both the hit category and the confidence level.

[0070] Step 4: Calculate the comprehensive risk score based on the business scenario weights; specifically including: Step 4.1: Based on the numbers of each hit category in the preliminary risk identification results, query the preset business scenario and risk weight mapping table. This table uses the risk category number and the business scenario number as a joint primary key, and the number of rows changes dynamically as the category set increases or decreases. Business scenarios include, but are not limited to, deployment environments such as healthcare, financial lending, food and daily necessities, and education and training.

[0071] The weight values ​​are derived from the historical proportion of violation cases in each scenario and the grading of regulatory penalties: the higher the proportion of a certain category in a certain scenario and the heavier the penalty, the greater the weight. All weights are dimensionless positive numbers between zero and one. Only the weights of the categories that are hit in the current campaign scenario are retrieved; categories that are not hit are not included in the calculation.

[0072] Step 4.2: Normalize the obtained weights: Divide the weight of each hit category by the sum of the weights of all hit categories to obtain the normalized business weight, which is dimensionless, and the sum of the normalized weights of all hit categories is always one. This ensures that regardless of the number of hit categories, the overall score falls within a fixed range, and the scores of different materials can be compared horizontally.

[0073] Step 4.3: Multiply the normalized business weights by the corresponding category's material-level confidence scores, then sum the results for all hit categories to obtain the comprehensive risk score: ; In the formula: N is the number of hit categories in step 3.3; c and d are the category summation subscripts; The business weight for category c is obtained in step 4.1. It is dimensionless and its fraction is the normalized business weight from step 4.2. R represents the material-level confidence level for category c, which is dimensionless and ranges from zero to one; R is the comprehensive risk score, which is dimensionless. Since the sum of the normalized weights is one and each confidence level does not exceed one, R falls within the range of zero to one.

[0074] Continuing the previous example: Suppose the current material is placed in a healthcare scenario, with false advertising and misleading consumption weights of 0.5 and 0.3 respectively, normalized to 0.625 and 0.375. Then R equals 0.625 multiplied by 0.72 plus 0.375 multiplied by 0.63, which is 0.45 plus 0.236, approximately 0.69. Since step 3.2 uses independent normalization for each category, the confidence levels do not dilute each other when multiple categories are hit simultaneously, and R can accurately reflect the risk accumulation.

[0075] Step 4.4: Determine the risk level. The scoring threshold is determined based on the quantile of the scoring distribution in the historical material review records, for example, using the 90th percentile and the 60th percentile, which in this embodiment are approximately 0.7 and 0.4 respectively: a score not lower than the first threshold is considered high risk, a score between the two thresholds is considered medium risk, and a score lower than the second threshold is considered low risk.

[0076] An additional direct pass rule for severe categories is established: when the confidence level of any material hitting a category is not lower than the preset severity threshold (e.g., 0.9), and that category belongs to the preset strong regulatory hazard classification, it is directly classified as high risk regardless of the score, compensating for the deficiency that a single extremely strong piece of evidence may be diluted by the average under the weighted average caliber. Low-risk materials are terminated early at this point, and their result in step four is the final output, not proceeding to steps five and six; high- and medium-risk materials, along with their score results, proceed to step five. In the previous example, the score was 0.69, and no category reached 0.9, so it was judged as medium risk.

[0077] Step 5: Perform cross-modal attribution decomposition on the comprehensive risk score; specifically including: Step 5.1: For the high- and medium-risk materials identified in Step 4, take the set of supporting anchor points for all hit categories and remove duplicates to form the target semantic anchor point set, denoted as M, where M is a positive integer not less than 1, determined by the actual violation clues hit. In the previous example, false advertising supporting anchor points one and two, and inducement to consume supporting anchor points two and four, after deduplication the target set consists of anchor points one, two, and four, and M equals 3.

[0078] Step 5.2: Perform baseline replacement perturbation on each target anchor point. First, calculate the baseline vector: take the arithmetic mean of the fusion vectors of all anchor points (including those not included in the target set) in the fusion feature sequence.

[0079] Subsequently, for a target anchor point, its fusion vector is replaced entirely with the baseline vector, while the remaining fusion vectors remain unchanged. The replaced sequence is then run again using the complete methods from steps three and four to obtain the post-disturbance comprehensive risk score for the anchor point after the replacement.

[0080] The mean baseline is used instead of zero because the zero vector is an off-distribution input. After passing through the classifier bias term and the normalization layer, it still produces non-zero activations, which will mix in components unrelated to evidence removal into the perturbation result. The mean vector is in the same distribution as the other fused vectors, and the perturbation change can more purely reflect the actual contribution of the anchor point evidence. The M target anchor points are replaced one by one and recalculated one by one to obtain M perturbation scores.

[0081] Previous calculations: After replacing anchor point one, the baseline confidence level for false advertising at the anchor point level is approximately 0.45, which is lower than the anchor point level threshold of 0.5. The confidence level for false advertising materials drops to 0.59 at anchor point two, which is lower than the category threshold of 0.6 and thus exits the hit. Only the category of inducement to consume is hit, with a weight normalized to 1, and the post-perturbation score is approximately 0.63. After replacing anchor point two, the confidence level for inducement to consume materials drops to 0.54 at anchor point four and exits the hit. Only the category of false advertising is hit, and the post-perturbation score is 0.72, which is higher than the original score. This indicates that the absence of this anchor point actually allows the category with higher confidence to monopolize the weight, and its contribution is counted as the absolute value of the change. After replacing anchor point four, the confidence levels for both types of materials remain unchanged, and the post-perturbation score is still 0.69, with a change of zero.

[0082] Step 5.3: Calculate the attribution contribution percentage of each target anchor point using the following formula: ; In the formula: M is the number of target anchor points; The subscript for the anchor point summation; R is the original comprehensive risk score from step 4.3, which is dimensionless; The post-perturbation score for the m-th target anchor point is the same as R in terms of caliber and dimension; the numerator is the absolute value of the difference between the two, representing the range of score change caused by the lack of evidence at that anchor point. The absolute value is taken so that both the increase and decrease in score are counted as contributions. The contribution percentage attribution is dimensionless, ranging from zero to one, and the sum of the percentages of all target anchor points is always one.

[0083] Boundary convention: When all denominator items are zero, meaning the absence of any target anchor point does not change the score and the evidence is completely redundant, the proportion of each anchor point is recorded as 1 / M, and the report will indicate evidence redundancy and no indistinguishable main contribution anchor points. In the previous example, the three change ranges were 0.06, 0.03, and 0, respectively, with a sum of 0.09, and proportions of approximately 0.67, 0.33, and 0.00.

[0084] Step 5.4: Assemble the contribution percentage, hit category number, material-level confidence level, and risk level in the order of anchor point numbers to generate a structured risk attribution report, divided into three layers: The first layer records the unique identifier, risk level, and overall score of the material. The second layer lists the contribution percentage, associated category number, and spatiotemporal location of the three modalities for each anchor point. Reviewers can directly jump to the corresponding segment of the material for verification based on the location. The third layer provides a natural language summary of which modalities and content segments contain the violation clues for anchor points whose contribution ratio exceeds the preset display threshold (e.g., 0.25), and which category of conclusions they mainly support.

[0085] In the previous example, anchor points one and two entered the third layer, while anchor point four was listed only in the second layer because its proportion was zero. The fact that its proportion was zero itself indicates that the evidence of this anchor point highly overlapped with that of anchor point two.

[0086] Step Six: Perform secondary verification and confidence correction on the recognition results based on cross-modal consistency; specifically including: Step 6.1: Based on the attribution report, anchor points whose contribution percentage exceeds the preset verification trigger threshold (e.g., 0.15) are selected as verification targets. Those whose contribution percentage does not reach the threshold have limited impact and are not verified. In the previous example, anchor point 1 (0.67) and anchor point 2 (0.33) are verified, while anchor point 4 (0.00) is skipped.

[0087] Step 6.2: Check the consistency of cross-modal evidence for each anchor point to be verified, and directly call the original similarity value retained in step 2.3.

[0088] For each modal combination of anchor point q, the arithmetic mean of the similarity of all feature pairs within the coverage area of ​​q under that combination is taken as the cross-verification strength of that combination. The similarity is not normalized and aggregated, and the value is independent of the number of coverage units of any modality. Therefore, the strength of the three combinations can be compared fairly, avoiding the problem of the strength being contaminated by the difference in coverage scale when using normalized weights as a metric.

[0089] Next, according to the variance metric, the consistency variance is calculated by dividing the sum of the squares of the differences between the strengths of the three combinations and their means by three. All quantities are dimensionless. When the variance is lower than the preset consistency threshold (e.g., 0.01), it indicates that the violation clues are balanced and mutually corroborated among the three modal combinations, and the conclusion is reliable. When it is not lower than the threshold, it indicates that the evidence is biased towards a particular modal combination, which may be due to one-sided coincidence. This anchor point is marked as a cross-modal inconsistency anchor point.

[0090] Previous calculations: The three combination strengths of anchor point one are 0.62 for text to image, 0.20 for text to audio, and 0.55 for image to audio, with a mean of approximately 0.457 and a variance of approximately 0.034, which is not lower than 0.01. The audio side evidence is obviously weak and is marked as inconsistent; The three strengths of anchor point two are 0.45, 0.42, and 0.47, with a mean of approximately 0.447 and a variance of approximately 0.0004, which is lower than the threshold and is considered consistent.

[0091] Step 6.3: Perform anchor-level confidence decay on inconsistent anchors. The decay only applies to the anchor-level confidence of the anchor itself and does not affect other anchors, thus avoiding the punishment of material-level conclusions by incorrect anchor-level marking.

[0092] The attenuation factor is linearly interpolated according to the degree of inconsistency: with a consistency threshold of 0.01 as the lower bound and a preset upper limit of variance (e.g., 0.05) as the upper bound, the factor decreases linearly from 0.7 to 0.3 as the variance increases. When the variance is not lower than the upper limit, it is directly taken as 0.3. The calculation method is 0.7 minus the difference between the variance and the lower bound, divided by the difference between the upper and lower bounds, and then multiplied by 0.4. In the previous example, the anchor point has a variance of 0.034. The factor is 0.7 minus 0.024, divided by 0.04, and then multiplied by 0.4, i.e., 0.7 minus 0.24, which gives 0.46. The anchor point confidence level of all categories of this anchor point is multiplied by 0.46: false advertising is corrected from 0.72 to 0.33, and induced consumption is corrected from 0.45 to 0.21. The confidence levels of consistent anchor points and unverified anchor points remain unchanged. After attenuation, the confidence level is still dimensionless and falls in the range of zero to one, consistent with the subsequent calculation method.

[0093] Step 6.4: Re-execute the two-level judgment in Step 3.3 and the scoring and grading in Step 4 with the corrected anchor-level confidence: In the previous example, only anchor two (0.59) remains as the supporting anchor for false advertising. The material-level confidence of 0.59 is below 0.6 and thus exits the hit; the material-level confidence of inducement to consume remains at 0.63 and remains in the hit; N' equals 1, the weight is normalized to 1, and the corrected score is 0.63, still between 0.4 and 0.7, maintaining a medium-risk level. If all categories exit the hit after correction, the score is recorded as zero and the level is low-risk.

[0094] Subsequently, the attribution report is updated synchronously according to the correction results: the first level of grades and scores are rewritten with the corrected values, and the third level is the anchor point that triggered the correction, with annotations on the confidence level, score, triggering reason, i.e., the consistency variance value and weak mode combination before and after the correction, so that the report is completely consistent with the final conclusion and no old values ​​that have been overturned are left behind.

[0095] The revised score, revised grade, and updated attribution report together constitute the final violation risk identification output of this embodiment: the score and grade answer how great the risk is, the attribution report answers which material segments and modal evidence the conclusions are derived from, and the consistency check ensures that the conclusions are based on stable evidence that corroborates each other across modalities.

[0096] The above description is merely an exemplary embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention specification and drawings under the technical concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for identifying violations of regulations in AI-generated advertising content, characterized in that, include: It receives AI-generated multimodal advertising materials, separates them into text stream, image stream, and audio stream, extracts semantic anchors of the three modalities, and generates structured material representations; Using semantic anchors as constraints, directional cross-modal joint feature fusion is performed on the structured material representation to obtain an anchor-guided fused feature sequence; The fusion feature sequence guided by the anchor point is input into the violation risk classifier, and the violation risk is independently identified for each category to obtain the preliminary risk identification result with anchor point positioning. Based on the business scenario weights, a comprehensive risk score is calculated and a risk level is determined for the preliminary risk identification results; The comprehensive risk score is subjected to cross-modal attribution decomposition to generate a risk attribution report; Based on cross-modal consistency, the identification results are subjected to secondary verification and confidence correction, and the corrected comprehensive risk score, corrected risk level and updated risk attribution report are output.

2. The method for identifying the violation risk of AI-generated advertising content according to claim 1, characterized in that, The extraction of three types of modal semantic anchors includes: The text stream is scanned according to the text rule set configured based on the preset set of violation risk categories. The sentence fragments that hit the rules are marked as text semantic anchors, and the anchor number, associated risk category number and character-level start and end positions are recorded. Text recognition is performed on the image stream. The text on the screen is compared with the text semantic anchors according to the bidirectional character overlap rate. When the preset standard is met, the anchor number of the corresponding text semantic anchor is used. The visual symbol detector independently establishes image semantic anchors for non-text visual elements according to the preset illegal visual symbol library. Anchors with the same associated risk category number and whose display time overlaps with the display time of the subtitle corresponding to the text semantic anchor are merged. If there are multiple candidates, the one with the longest overlap time is merged. The audio stream is converted from speech to text to obtain narration text. The anchor point number of the text semantic anchor point is used according to the bidirectional word overlap rate. The audio feature discriminator independently establishes audio semantic anchor points for suspicious acoustic segments based on spectral features and processes them according to the same merging rules. The merged three types of modal semantic anchors are mapped to a unified time axis, and the structured material representation is output.

3. The method for identifying the violation risk of AI-generated advertising content according to claim 2, characterized in that, Perform targeted cross-modal joint feature fusion on the structured material representation, including: The three contents in the structured material representation are vectorized respectively and arranged into three initial feature sequences; Generate a three-way mask vector for each semantic anchor point one by one, and simultaneously select one from the masks of multiple semantic anchor points that fall into the same feature unit. The similarity between feature units is calculated and divided by a preset scaling factor before an exponential transformation is performed. Only feature units that are simultaneously activated by the mask of the same semantic anchor point are allowed to interact. After row normalization, the cross-modal alignment weight with anchor point index is obtained. When a certain modality combination has no covered unit on the semantic anchor point, all cross-modal alignment weights under the modality combination are set to zero. Each of the three modal pooling vectors is concatenated with the enhancement terms in the three directions one by one according to the anchor point. The resulting fusion vectors of each semantic anchor point are obtained by linear mapping compression and arranged in the order of the anchor point numbers to form the fusion feature sequence guided by the anchor points.

4. The method for identifying the violation risk of AI-generated advertising content according to claim 3, characterized in that, The separate identification of violation risks by category includes: The fusion vector of each semantic anchor in the fusion feature sequence guided by the anchor is input into the violation risk classifier to obtain the original discriminant value of each semantic anchor in each risk category; Each original discriminant value is normalized independently to obtain the anchor-level confidence score, and the anchor-level confidence scores of each category are not diluted. For each risk category, all semantic anchors with anchor-level confidence scores not lower than the preset anchor-level threshold are included in the support anchor set. The highest anchor-level confidence score under the category is used as the material-level confidence score. Categories with material-level confidence scores higher than the preset category judgment threshold are included in the initial hit risk set. The category judgment threshold is set higher than the anchor-level threshold. Integrate the three types of spatiotemporal positions of each support anchor point, and output the preliminary risk identification results with anchor point positioning.

5. The method for identifying the violation risk of AI-generated advertising content according to claim 4, characterized in that, The calculation of the comprehensive risk score and the determination of the risk level include: Based on the risk category number of each hit category in the preliminary hit risk set, query the preset business scenario and risk weight mapping table to obtain the business weight coefficient of each hit category in the current deployment scenario; The normalized business weight is obtained by dividing the business weight coefficient of each hit category by the sum of the business weight coefficients of all hit categories. The normalized business weights are multiplied by the corresponding category's material-level confidence scores, and then summed over all hit categories to obtain the comprehensive risk score. The risk level is determined by comparing the comprehensive risk score with the score threshold defined by the quantile of the score distribution based on the historical material review records. If the confidence level of any material that hits a category is not lower than the preset severity threshold and belongs to the preset strong regulatory hazard level, it is directly defined as high risk. Materials with a low risk level are terminated in advance, and the comprehensive risk score and risk level are used as the final output.

6. The method for identifying the violation risk of AI-generated advertising content according to claim 5, characterized in that, The cross-modal attribution decomposition includes: Take the union of all support anchor points of the hit categories and remove duplicates to form the target semantic anchor point set; Using the arithmetic mean of the fusion vectors of all semantic anchors as the baseline vector, the fusion vector of each target semantic anchor is replaced one by one with the baseline vector. The entire process of independently identifying violation risks by category, calculating the comprehensive risk score and defining the risk level is re-executed with the fusion feature sequence guided by the replaced anchors, so as to obtain the post-perturbation comprehensive risk score of each target semantic anchor. The attribution contribution ratio of each target semantic anchor is obtained by dividing the absolute value of the difference between the comprehensive risk score and each post-disturbance comprehensive risk score by the sum of the differences of all target semantic anchors. When all differences are zero, the attribution contribution ratio of each target semantic anchor is recorded as the reciprocal of the number of target semantic anchors, and evidence redundancy is marked.

7. The method for identifying the violation risk of AI-generated advertising content according to claim 6, characterized in that, The hierarchical generation of the risk attribution report includes: The first layer records the unique identifier, risk level, and comprehensive risk score of the material. The second layer lists the attribution contribution percentage, associated risk category number, and spatiotemporal location of each target semantic anchor point according to the anchor point number. The third layer identifies target semantic anchors whose attribution contribution exceeds a preset display threshold, and uses natural language to summarize the modalities and content segments where violation clues appear, as well as the risk category conclusions they support.

8. The method for identifying the violation risk of AI-generated advertising content according to claim 7, characterized in that, The secondary verification includes: Select target semantic anchors whose attribution contribution ratio exceeds the preset verification trigger threshold as objects to be verified. For each object to be verified, the arithmetic mean of the original similarities of all feature pairs within the coverage of each modality combination is taken as the cross-verification strength of the modality combination. The original similarities are not normalized and aggregated and are independent of the number of coverage units of any modality. The consistency variance is calculated based on the interaction verification strength of the three modal combinations. When the consistency variance is not lower than a preset consistency threshold, the corresponding object to be verified is marked as a cross-modal inconsistency anchor point.

9. The method for identifying the violation risk of AI-generated advertising content according to claim 8, characterized in that, The confidence level correction includes: Using the consistency threshold and the preset upper limit of variance as interpolation boundaries, the attenuation factor is obtained by linear interpolation according to the consistency variance. The larger the variance, the smaller the attenuation factor. The anchor-level confidence scores of all risk categories of the cross-modal inconsistency anchor itself are multiplied by the attenuation factor, while the anchor-level confidence scores of the remaining semantic anchors remain unchanged. The two-level judgment and scoring classification are re-executed with the revised anchor-level confidence level to obtain the revised comprehensive risk score and the revised risk level. The risk attribution report is updated synchronously with the first-level level and score record based on the correction results. The cross-modal inconsistency anchor points that trigger the correction are annotated with the confidence level, score, consistency variance and weak mode combination before and after the correction, and the updated risk attribution report is output.

10. A system for identifying violations of regulations in AI-generated advertising content, characterized in that, The method for identifying violations of AI-generated advertising content as described in any one of claims 1-9 includes: The receiving unit is used to receive AI-generated multimodal advertising materials, separate them into text stream, image stream and audio stream, extract semantic anchors of the three modalities, and generate structured material representations; The fusion unit is used to perform directional cross-modal joint feature fusion on the structured material representation with semantic anchors as constraints, so as to obtain an anchor-guided fusion feature sequence; The classification unit is used to input the fusion feature sequence guided by the anchor point into the violation risk classifier, independently identify the violation risk by category, and obtain the preliminary risk identification result with anchor point positioning; The scoring unit is used to calculate a comprehensive risk score and determine the risk level based on the business scenario weights of the preliminary risk identification results; The attribution unit is used to perform cross-modal attribution decomposition on the comprehensive risk score and generate a risk attribution report; The verification unit is used to perform secondary verification and confidence correction on the identification results based on cross-modal consistency, and output the corrected comprehensive risk score, the corrected risk level, and the updated risk attribution report.

Images