Advertisement label labeling method, electronic equipment and storage medium

Abstract

The invention relates to the technical field of advertisement label labeling methods, in particular to an advertisement label labeling method, electronic equipment and a storage medium, and the method specifically comprises the following steps: step 1, carrying out stage division and core target definition on an advertisement life cycle, including life cycle stage division and stage core target quantification; step 2, carrying out cross-modal advertisement information collection and semantic initial extraction, including multi-modal information collection, single-modal semantic initial extraction, visual semantic extraction and audio semantic extraction; step 3, performing cross-modal semantic fusion and stage label generation, including inter-modal semantic association degree calculation and stage fusion label generation; and step 7, performing life cycle stage switching and label updating, including stage switching detection and label updating process setting. According to the invention, through life cycle stage division and core target quantification, the problem that static annotation is disjointed with an advertisement stage is solved.

Description

Technical Field

[0001] This invention relates to the technical field of labeling methods for advertising labels, and in particular to a labeling method for advertising labels, an electronic device, and a storage medium. Background Technology

[0002] In the current advertising ecosystem, ad tags serve as a core bridge connecting ad content and user needs, and their accuracy directly impacts ad matching efficiency and user experience.

[0003] Current advertising tagging technologies mostly focus on static tagging in the initial stage of ad launch, lacking dynamic adaptation throughout the entire ad lifecycle, and have significant shortcomings in semantic fusion of multimodal content (text, images, videos, and audio). Among the existing mainstream methods, text-driven tagging only extracts keywords from the ad copy to generate tags, failing to interpret implicit information in visual and audio contexts. For example, visual features such as "fabric texture" and "wearing scenarios" in clothing ads are difficult to accurately capture through text. Visual-driven tagging relies on image recognition technology to generate tags, but it is prone to "semantic isolation" problems, such as recognizing the "red" tag but failing to associate it with the marketing scenario semantics of "holiday promotion." Summary of the Invention

[0004] In response to the technical problems mentioned in the background art, the present invention provides a labeling method for advertising labels, an electronic device, and a storage medium.

[0005] The technical solution adopted in this invention is: a labeling method for advertising tags, specifically including the following steps:

[0006] Step 1: Divide the advertising lifecycle into stages and define core objectives, including dividing the lifecycle into stages and quantifying the core objectives for each stage;

[0007] Step 2: Perform cross-modal advertising information collection and initial semantic extraction, including multimodal information collection, initial semantic extraction of a single modality, visual semantic extraction, and audio semantic extraction;

[0008] Step 3: Perform cross-modal semantic fusion and stage label generation, including calculating the semantic correlation between modalities and generating stage fusion labels;

[0009] Step 4: Construct an industry-specific semantic dictionary and calibrate tags, including constructing an industry-specific dictionary, calculating the industry suitability of tags, and performing tag calibration operations;

[0010] Step 5: Perform compliance and privacy risk label filtering, including compliance risk score calculation, privacy risk detection, and filtering rule establishment;

[0011] Step 6: Conduct campaign performance feedback and tag weight iteration, including performance metric collection and tag weight iteration;

[0012] Step 7: Perform lifecycle stage switching and tag updates, including stage switching detection and tag update process settings.

[0013] In one embodiment, in step 1:

[0014] The lifecycle stages are divided as follows;

[0015] Based on the advertising campaign cycle and budget allocation, the advertising lifecycle is divided into three stages: the pre-launch phase, the peak phase, and the decline phase, denoted by stage S. lc It indicates that the preheating period S lc =1, outbreak period S lc =2, Decline period S lc =3;

[0016] The core objectives for this phase are quantified as follows:

[0017] Define and quantify the core objectives for each stage, using the formula G. stage =a×G exp +b×G clk +c×G conv ;

[0018] Among them: G exp For exposure target weight; G clk For click target weight; G conv To transform the target weights; a+b+c=1, G stage The value ranges from 0 to 1, representing the weight allocation of the core objectives at this stage.

[0019] In one embodiment, in step 2:

[0020] Multimodal information acquisition is as follows;

[0021] Collect text information from advertisements, including title (T), detail page copy (D), advertising slogan (S); visual information, including main image (V1), scene image (V2), video keyframes (V3); and audio information, including advertisement voice-over (A1) and background music style (A2), to form a multimodal information set Info. multi ={T,D,S,V1,V2,V3,A1,A2};

[0022] The initial extraction of single-modal semantics is as follows;

[0023] Text semantic extraction: A BERT pre-trained model is used to extract keywords from the text, and the keyword importance W is calculated. txt The formula is Among them TF kIDF represents the word frequency of keyword k in the text. k W represents the inverse document frequency of keyword k, n is the total number of extracted keywords, and W is the inverse document frequency of keyword k. txt The value range is 0-1, and W is selected. txt The top 5 keywords are used as the initial text tags L t xt-init;

[0024] Visual semantic extraction is as follows;

[0025] The YOLOv8 object detection model was used to identify visual elements, and the element confidence score C was calculated. vis The formula is

[0026] IoU is the intersection-union ratio of the detection boxes, Conf is the model output information, and C is the C value. vis The value range is 0-1, and C is selected. vis Elements with a value ≥0.7 are used as the initial visual label L. vis-init ;

[0027] The audio semantics extraction is as follows;

[0028] The voiceover A1 is converted into text using a speech recognition model, and the background music style A2 is determined using an audio sentiment analysis model to generate the initial audio tag L. aud-init .

[0029] In one embodiment, in step 3:

[0030] The semantic association degree between modalities is calculated as follows;

[0031] Sim calculates the semantic association between text, visual, and audio tags. cross The formula is

[0032]

[0033] Where: Sim txt-vis For semantic similarity between text and visual labels; Sim vis-aud For semantic similarity between visual and audio tags; Sim cross The value ranges from 0 to 1, with higher values ​​indicating stronger semantic consistency of multimodal tags.

[0034] The stage fusion tags are generated as follows;

[0035] Combining the core objectives of the life cycle stage G stage Sim with cross-modal semantic association cross Generation stage fusion label L stage The formula is L stage ={L|W l ×Simcross ×G stage ≥0.5}, where W l Assuming weights for the initial labels of a single modality, labels with a comprehensive score ≥ 0.5 are selected as the core labels for this stage.

[0036] In one embodiment, in step 4:

[0037] The industry-specific dictionary is constructed as follows;

[0038] For different industries, we collect industry-specific terminology, product characteristic terms, and compliance keywords to build an industry-specific semantic dictionary, Dict. ind It includes three fields: terminology definition, semantic weight, and compliance level;

[0039] The industry compatibility of the label is calculated as follows;

[0040] Computational stage fusion label L stage Dict Industry Dictionary ind Adaptability F ind The formula is

[0041] Where: N match For L stage Number of matches with dictionary terms; W ind The semantic weight of the matching terms; m is the total number of terms in the dictionary; the value ranges from 0 to 1, and the higher the value, the stronger the industry adaptability of the tag;

[0042] The label calibration procedure is as follows;

[0043] If F ind ≥0.7, keep L directly. stage If 0.4≤F ind If F < 0.7, supplement the dictionary with terms with high semantic weight; if F ind If the value is less than 0.4, repeat steps 2-3 to optimize the initial tag extraction rules based on the industry dictionary.

[0044] In one embodiment, in step 5:

[0045] The compliance risk score is calculated as follows;

[0046] Building a compliance risk thesaurus (Risk) com Calculate the label compliance risk score R com The formula is Where Sim(l,R) com-i W represents the semantic similarity between the tag and the risk word. risk For risk word weights, R com The value range is 0-1.

[0047] The following privacy risk assessments were conducted.

[0048] Building a privacy-sensitive word library Risk pri Calculate the privacy risk of the label R pri The formula is Where N pri-match Total represents the number of matches between tags and the privacy thesaurus. l R represents the total number of labels. pri The value range is 0-1;

[0049] The filtering rules are as follows;

[0050] If R com >0.3 or R pri If >0, filter the tag directly; if R com ≤0.3 and R pri =0, retain the label and record the compliance score for subsequent performance tracking.

[0051] In one embodiment, in step 6:

[0052] The performance indicators were collected as follows;

[0053] Real-time collection of campaign performance data corresponding to tags: Tag exposure (Imp) l Click count l Conversion amount Conv l Calculate the label effect score S e ff, the formula is S e ff = α × CTR l +β×CVR l ;

[0054] in: Tag click-through rate; Tag conversion rate; α = 0.4, β = 0.6; S eff The value ranges from 0 to 1, representing the actual marketing value of the tag;

[0055] The label weight iteration is as follows;

[0056] According to S eff Update tag weight W l -new, the formula is W l -new=W l-old ×(1+γ×S eff ), where W l -old represents the original weights, γ = 0.2 represents the iteration coefficients, and W... l-new The value range is 0-1.5; if S eff If W < 0.2, then reduce the weight W. l -new=Wl-old ×0.8.

[0057] In one embodiment, in step 7:

[0058] Phase switching detection is as follows;

[0059] When the advertising campaign reaches the lifecycle stage threshold, or the core objectives of the stage are achieved... Trigger phase switching, update phase identifier S lc .

[0060] The tag update process is as follows;

[0061] After the phase switch, re-execute steps 2-6, based on the new phase's G. stage Industry dictionary and latest performance data are used to generate new stage labels L. s `tag-new` replaces the original tag; at the same time, it retains the historical tag record. h istory={L stage1 ,L stage2 ,L stage3} is used for retrospective analysis of advertising performance.

[0062] The beneficial effects of this invention are as follows: Compared with the prior art, this invention solves the problem of the disconnect between static labeling and advertising stages by dividing the lifecycle into stages and quantifying core objectives, making the tags adapt to the promotion focus of each stage: preheating, peak, and decline; it compensates for the shortcomings of single-modal labeling by cross-modal information collection and initial semantic extraction; it achieves multimodal semantic association by leveraging cross-modal semantic fusion and stage tag generation, ensuring that the tags not only meet the stage objectives but also fully reflect the core value of the advertisement, solving the problem of semantic isolation of multimodal tags; it improves the accuracy of vertical industry labeling by constructing an industry-specific semantic dictionary and calibrating the tags, avoiding misjudgment of professional terms and meeting the compliance and precise targeting needs of vertical industries; it filters for compliance and privacy risks by filtering out violations in advance, protecting user privacy and solving the problem of delayed compliance review; it quantifies the actual marketing value of tags and dynamically adjusts the weights based on the feedback of the campaign performance and the iteration of tag weights, prioritizing the matching of high-value tags and solving the problem of tags without performance verification and fixed weights; and it ensures that tags are dynamically adjusted with the advertising stage by switching the lifecycle stage and updating the tags, while retaining historical tags for performance backtracking, solving the problem of tags being fixed for a long time and unable to adapt to stage changes. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the present invention; Detailed Implementation

[0064] In the description of this invention, it should be noted that the terms "front", "up", "down", "left", "right", "vertical", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0065] To address the problems existing in the background art, this application proposes the following technical solution: a labeling method for advertising labels, an electronic device, and a storage medium;

[0066] Specifically, the following steps are included:

[0067] Step 1: Defining the advertising lifecycle stages and core objectives;

[0068] Lifecycle phase division;

[0069] Based on the advertising campaign cycle and budget allocation, the advertising lifecycle is divided into three stages: the pre-launch phase (1-7 days before launch, 10% of the budget), the peak phase (8-30 days after launch, 60% of the budget), and the decline phase (after 31 days after launch, 30% of the budget). These stages are labeled S. lc Indicates (preheating period S) lc =1, outbreak period S lc =2, Decline period S lc =3).

[0070] Quantifying the core objectives of each stage;

[0071] Define and quantify the core objectives for each stage, using the formula G. stage =a×G exp +b×G clk +c×G conv ;

[0072] Among them: G exp To determine the exposure target weight (focusing on exposure during the pre-heating period, G) exp =0.6; G during the outbreak period exp =0.3; G during the decline period exp =0.2); G clk Click target weight (preheating period G) clk =0.3; During the burst period, focus on clicks, G clk =0.5; G during the decline period clk =0.3); G conv To transform the target weight (preheating period G) conv =0.1; G during the outbreak period conv =0.2; the decline phase focuses on transformation, G conv=0.5); a+b+c=1, G stage The value ranges from 0 to 1, representing the weight allocation of the core objectives at this stage.

[0073] The above technical solution is explained as follows: By clearly defining the three stages of advertising—pre-launch, peak, and decline—and quantifying the core objectives of each stage, the problem of static labeling being disconnected from the advertising lifecycle in existing technologies is effectively solved. Previously, labeling was only completed when the ad went live, failing to adapt to the different promotional focuses at each stage. This step provides stage-specific guidance for subsequent label generation. For example, the pre-launch stage focuses on exposure goals, guiding the generation of "New Product Preview" tags; the peak stage focuses on click goals, driving the generation of "Limited-Time Discount" tags, ensuring that tags highly align with the promotional needs of each stage. Simultaneously, the binding of budget allocation to stage objectives makes labeling more aligned with advertising resource allocation, avoiding resource waste and laying a stage-specific foundation for accurate subsequent label matching, thus improving the targeting and efficiency of the entire advertising lifecycle.

[0074] Step 2: Cross-modal advertising information collection and initial semantic extraction;

[0075] Multimodal information acquisition;

[0076] Collect text information (title T, detail page copy D, advertising slogan S), visual information (main image V1, scene image V2, video keyframes V3), and audio information (ad voice-over A1, background music style A2) from the advertisement to form a multimodal information set Info. multi ={T,D,S,V1,V2,V3,A1,A2}.

[0077] Initial extraction of single-modal semantics;

[0078] Text semantic extraction: A BERT pre-trained model is used to extract keywords from the text, and the keyword importance W is calculated. txt The formula is Among them TF k IDF represents the word frequency of keyword k in the text. k W represents the inverse document frequency of keyword k, n is the total number of extracted keywords, and W is the inverse document frequency of keyword k. txt The value range is 0-1, and W is selected. txt The top 5 keywords are used as the initial text tags L t xt-init.

[0079] Visual semantic extraction;

[0080] The YOLOv8 object detection model is used to identify visual elements (such as products, scenes, and people), and the element confidence score C is calculated. vis The formula is

[0081] IoU is the intersection-union ratio of the detection boxes, Conf is the model output information, and C is the C value. vis The value range is 0-1, and C is selected. vis Elements with a value ≥0.7 are used as the initial visual label L. vis-init .

[0082] Audio semantic extraction;

[0083] The voiceover A1 is converted into text using a speech recognition model, and the background music style A2 (such as "cheerful" or "calm") is determined using an audio sentiment analysis model to generate the initial audio tag L. aud-init .

[0084] The above technical solution is explained as follows: By collecting multimodal information from text, vision, and audio, and combining it with professional models and formulas, initial semantic extraction is achieved, significantly improving the comprehensiveness and accuracy of initial tag extraction. In text semantic extraction, core keywords are accurately selected through word frequency and inverse document frequency normalization calculations, avoiding redundancy or omission of text keywords. In visual semantic extraction, visual tags are ensured to be authentic and reliable, such as accurately identifying implicit visual features like "fabric texture" in clothing advertisements. Audio semantic extraction supplements audio dimension information, such as the "cheerful" background music style tag, filling the gap in traditional methods that ignore audio semantics. The comprehensive collection and accurate initial extraction of multimodal information provide high-quality basic data for subsequent cross-modal semantic fusion, avoiding the fragmentation problem of single-modal tags, and making the tags more reflective of the multi-dimensional core information of the advertisement.

[0085] Step 3: Cross-modal semantic fusion and stage label generation;

[0086] Calculation of semantic association between modalities;

[0087] Sim calculates the semantic association between text, visual, and audio tags. cross The formula is

[0088]

[0089] Where: Sim txt-vis The semantic similarity between text and visual labels is calculated using the Word2Vec model.

[0090] Sim vis-aud The semantic similarity between visual and audio tags (e.g., the similarity between "cheerful" audio and "festive scene" visual);

[0091] Sim cross The value ranges from 0 to 1, with higher values ​​indicating stronger semantic consistency of multimodal tags.

[0092] Stage fusion tag generation;

[0093] Combining the core objectives of the life cycle stage G stage Sim with cross-modal semantic association cross Generation stage fusion label L stage The formula is L stage ={L|W l ×Sim cross ×G stage ≥0.5}, where W l Weights of the initial label for a single modality (text W) l =0.4, visual W l =0.4, audio W l =0.2), and select tags with a comprehensive score ≥0.5 as the core tags for this stage.

[0094] The above technical solution is explained as follows: It achieves deep fusion of multimodal semantics and accurate generation of stage tags, solving the problem of isolated semantics and disconnection from stage goals in existing technologies. The semantic relevance of text, visual, and audio tags is calculated using formulas. For example, "red dress" (text), "festival party scene" (visual), and "upbeat background music" (audio) are associated and fused to avoid tags being independent. Tags are selected by combining the core stage goals with semantic relevance, ensuring that the generated tags not only conform to the stage's promotional focus but also possess multimodal semantic consistency. For example, during the peak period, a fused tag "festival party red dress limited-time discount" is generated, which not only matches the click target but also fully presents the core value of the advertisement.

[0095] Step 4: Construction of an industry-specific semantic dictionary and tag calibration;

[0096] Industry-specific dictionary construction;

[0097] For different industries (such as healthcare, finance, and e-commerce), we collect industry-specific terminology, product characteristic terms, and compliance keywords to build an industry-specific semantic dictionary, Dict. ind It includes three fields: term definition, semantic weight, and compliance level (e.g., in the medical industry dictionary, the semantic weight of "prescription drug" is 0.8, and the compliance level is "high").

[0098] Label industry compatibility calculation;

[0099] Computational stage fusion label L stage Dict Industry Dictionary ind Adaptability F ind The formula is

[0100] Where: N match For L stage Number of matches with dictionary terms; W indThe semantic weight of the matching terms; m is the total number of terms in the dictionary; the value ranges from 0 to 1, and the higher the value, the stronger the industry adaptability of the tag.

[0101] Label calibration operation;

[0102] If F ind ≥0.7 (high compatibility), directly retain L stage If 0.4≤F ind <0.7 (moderate fit), supplement the dictionary with high semantic weight terms (e.g., supplement financial advertisements with "risk level R2"); if F ind If the score is less than 0.4 (low compatibility), repeat steps 2-3 to optimize the initial tag extraction rules based on the industry dictionary.

[0103] The above technical solution is explained as follows: Addressing the pain points of vertical industry labeling, an industry-specific semantic dictionary is constructed, and by calculating the fit, the accuracy of vertical industry advertising labeling is significantly improved. Traditional general semantic models are prone to misjudging professional terms in fields such as healthcare and finance. The industry-specific semantic dictionary contains information such as professional term definitions and semantic weights, such as the medical industry term "prescription drug" and its semantic weight settings. The fit between tags and the industry dictionary is quantified, and tags are retained, supplemented, or re-extracted based on the fit, such as adding the professional tag "risk level R2" to financial advertisements. Tag calibration ensures that the generated tags meet industry professional characteristics and compliance requirements, avoiding inaccurate tags in professional fields. This meets the needs of accurate labeling and compliant placement of advertisements in vertical industries such as healthcare and finance, improving the professionalism and reliability of vertical industry advertising tags, and providing a compliant and professional tag foundation for subsequent compliance filtering and performance feedback.

[0104] Step 5: Compliance and privacy risk label filtering;

[0105] Compliance risk score calculation;

[0106] Building a compliance risk thesaurus (Risk) com (Including false advertising and prohibited words), calculate the label compliance risk score R. com The formula is Where Sim(l,R) com-i W represents the semantic similarity between the tag and the risk word. risk Weighting of risky keywords (high-risk W) risk =1, medium risk W risk =0.6, low risk W risk =0.3), R com The value range is 0-1.

[0107] Privacy risk monitoring;

[0108] Building a privacy-sensitive word library Risk pri(Including "phone number", "location", etc.), calculating tag privacy risk R pri The formula is Where N pri-match Total represents the number of matches between tags and the privacy thesaurus. l R represents the total number of labels. pri The value range is 0-1.

[0109] Filtering rules;

[0110] If R com >0.3 or R pri If >0, filter the tag directly; if R com ≤0.3 and R pri =0, retain the label and record the compliance score for subsequent performance tracking.

[0111] The above technical solution is explained as follows: Clear filtering rules enable precise early filtering of violation and privacy risk tags, addressing the issues of lagging compliance review and insufficient privacy protection in existing technologies. It quantifies tag compliance risks, such as identifying "false advertising" tags to prevent penalties for advertising violations; it detects privacy risks, such as filtering privacy-related tags like "phone number" and "location," complying with data security regulations and protecting user privacy. Clear filtering rules ensure that all high-risk tags are removed, while low-risk compliant tags are retained, controlling compliance and privacy risks at the tagging stage and avoiding traffic waste and compliance risks during subsequent advertising campaigns due to tag issues.

[0112] Step 6: Feedback on campaign performance and iterative evaluation of tag weights;

[0113] Performance metrics collection;

[0114] Real-time collection of campaign performance data corresponding to tags: Tag exposure (Imp) l Click count l Conversion amount Conv l Calculate the label effect score S e ff, the formula is S e ff = α × CTR l +β×CVR l ;

[0115] in: (Tag click-through rate); (Tag conversion rate); α = 0.4, β = 0.6 (conversion weight is higher than click); S eff The value ranges from 0 to 1, representing the actual marketing value of the tag.

[0116] Tag weight iteration;

[0117] According to Seff Update tag weight W l -new, the formula is W l -new=W l-old ×(1+γ×S eff ), where W l -old represents the original weights, γ = 0.2 represents the iteration coefficients, and W... l-new The value range is 0-1.5 (to avoid excessive weighting leading to label fixation). If S eff If W < 0.2, then reduce the weight W. l -new=W l-old ×0.8.

[0118] The above technical solution is explained as follows: By constructing a campaign performance feedback mechanism, the value of tags is quantified and their weights are dynamically iterated, solving the problems of existing technologies where tags lack performance verification and have fixed weights. Combining click-through rate (CTR) and conversion rate, the actual marketing value of tags is objectively quantified, avoiding a disconnect between "subjective labeling" and actual results, such as accurately identifying the high marketing value of the "high cost-performance" tag. Tag weights are dynamically adjusted based on performance scores; high-value tags receive increased weight and are prioritized in subsequent matching, while low-value tags have decreased weight, avoiding resource waste. For example, if the "discount" tag has a high CTR and conversion rate, its weight gradually increases, enhancing its priority in ad matching and ensuring that tags always closely align with user needs.

[0119] Step 7: Lifecycle stage switching and tag update;

[0120] Phase transition detection;

[0121] When the advertising campaign reaches the lifecycle stage threshold (e.g., the 7-day pre-launch period ends), or the core objectives of the stage are achieved... (Actual target exceeded expectation by 2096), triggering a phase switch and updating phase identifier S. lc .

[0122] Tag update process;

[0123] After the phase switch, re-execute steps 2-6, based on the new phase's G. stage Industry dictionary and latest performance data are used to generate new stage labels L. s `tag-new` replaces the original tag; at the same time, it retains the historical tag record. h istory={L stage1 ,L stage2 ,L stage3} is used for retrospective analysis of advertising performance.

[0124] The above technical solution is explained as follows: Through a phase-switching detection mechanism and a tag update process, tags are dynamically adjusted throughout the advertising lifecycle, solving the problem of existing technologies where tags remain fixed for extended periods and cannot adapt to phase changes. Phase-switching detection is based on the campaign duration and target achievement rate. For example, a switch is triggered when the 7-day pre-launch period ends or the target exceeds expectations by 20%, responding promptly to changes in the advertising phase. The tag update process re-executes the previous steps to generate new phase tags, such as updating to the "Clearance Sale" tag during the decline phase to align with conversion goals, while retaining historical tag records for performance review. Dynamic updates ensure that tags always align with the promotional needs of each phase, avoiding the misalignment of using the "New Product Preview" tag during the advertising decline phase. Historical tag records provide advertisers with a basis for performance analysis; by comparing the performance of tags at different phases, subsequent promotional strategies can be optimized.

[0125] In summary: First, this embodiment achieves several advantages. First, by dividing the advertising into pre-launch, peak, and decline phases and quantifying core objectives, the tags are dynamically adjusted according to the advertising stage. For example, during the pre-launch phase, tags like "New Product Preview" and "Early Bird Reservation" are generated; during the peak phase, tags like "Limited-Time Discount" and "Discount for Purchases Over a Certain Amount" are generated; and during the decline phase, tags like "Clearance Sale" and "Last 3 Days" are generated. This ensures that the tags always align with the promotional focus of each stage, improving the matching efficiency of the advertisement at different stages. Second, it achieves deep integration of cross-modal semantics. By calculating the semantic relevance of text, visual, and audio tags, it avoids the isolation of multimodal information. For example, the tags "red dress" (text), "festival party scene" (visual), and "upbeat background music" (audio) in a clothing advertisement can be integrated to generate the tag "festival party red dress," fully reflecting the core value of the advertisement and solving the fragmentation problem of traditional tagging methods. Third, it improves the accuracy of vertical industry labeling. By constructing an industry-specific semantic dictionary and calculating adaptability, it provides accurate labeling for advertisements in professional fields such as healthcare and finance. For example, medical advertisements can be accurately labeled with "prescription." The platform employs professional terminology such as "medicine" and "symptomatic treatment" to avoid misjudgments by general models and meet the compliance and precise targeting needs of vertical industry advertising. Fourth, it establishes a closed-loop feedback mechanism, collecting tag performance data and iterating weights to prioritize high-click-through-rate and high-conversion-rate tags while eliminating low-value tags. For example, if the "high cost-performance" tag consistently achieves a CTR and CVR above the average, its weight will gradually increase, leading to priority recommendations in subsequent matching, ensuring a high degree of alignment between tags and user needs. Fifth, it strengthens compliance and privacy protection by quantifying compliance and privacy risk values, filtering out illegal and sensitive tags in advance to prevent advertising from facing compliance penalties due to tag issues, while protecting user privacy and increasing user trust in the platform. Sixth, it provides data support for optimizing advertising strategies. Historical tag records allow for reviewing tag performance at different stages, helping advertisers adjust their promotion strategies. For example, analysis reveals that the "discount for purchases over a certain amount" tag has the best conversion rate during peak periods, allowing for increased advertising efforts related to this tag to achieve efficient use of advertising resources.

[0126] Although embodiments of the invention have been shown and described, the scope of the invention will be defined by the appended claims and their equivalents by those skilled in the art.

Claims

1. A method for labeling advertising tags, characterized in that, Specifically, the following steps are included: Step 1: Divide the advertising lifecycle into stages and define core objectives, including dividing the lifecycle into stages and quantifying the core objectives for each stage; Step 2: Perform cross-modal advertising information collection and initial semantic extraction, including multimodal information collection, initial semantic extraction of a single modality, visual semantic extraction, and audio semantic extraction; Step 3: Perform cross-modal semantic fusion and stage label generation, including calculating the semantic correlation between modalities and generating stage fusion labels; Step 4: Construct an industry-specific semantic dictionary and calibrate tags, including constructing an industry-specific dictionary, calculating the industry suitability of tags, and performing tag calibration operations; Step 5: Perform compliance and privacy risk label filtering, including compliance risk score calculation, privacy risk detection, and filtering rule establishment; Step 6: Conduct campaign performance feedback and tag weight iteration, including performance metric collection and tag weight iteration; Step 7: Perform lifecycle stage switching and tag updates, including stage switching detection and tag update process settings.

2. The labeling method for advertising labels according to claim 1, characterized in that, In step 1: The lifecycle stages are divided as follows; Based on the advertising campaign cycle and budget allocation, the advertising lifecycle is divided into three stages: the pre-launch phase, the peak phase, and the decline phase, denoted by stage S. lc It indicates that the preheating period S lc =1, outbreak period S lc =2, Decline period S lc =3; The core objectives for this phase are quantified as follows; Define and quantify the core objectives for each stage, using the formula G. stage =a×G exp +b×G clk +c×G conv ; Among them: G exp For exposure target weight; G clk For click target weight; G conv To transform the target weights; a+b+c=1, G stage The value ranges from 0 to 1, representing the weight allocation of the core objectives at this stage.

3. The labeling method for an advertising label according to claim 2, characterized in that, In step 2: Multimodal information collection is as follows; Collect text information from advertisements, including title (T), detail page copy (D), advertising slogan (S); visual information, including main image (V1), scene image (V2), video keyframes (V3); and audio information, including advertisement voice-over (A1) and background music style (A2), to form a multimodal information set Info. multi ={T,D,S,V1,V2,V3,A1,A2}; The initial extraction of single-modal semantics is as follows; Text semantic extraction: A BERT pre-trained model is used to extract keywords from the text, and the keyword importance W is calculated. txt The formula is Among them TF k IDF represents the word frequency of keyword k in the text. k W represents the inverse document frequency of keyword k, n is the total number of extracted keywords, and W is the inverse document frequency of keyword k. txt The value range is 0-1, and W is selected. txt The top 5 keywords are used as the initial text tags L t xt-init; Visual semantic extraction is as follows; The YOLOv8 object detection model was used to identify visual elements, and the element confidence score C was calculated. vis The formula is IoU is the intersection-union ratio of the detection boxes, Conf is the model output information, and C is the C value. vis The value range is 0-1, and C is selected. vis Elements with a value ≥0.7 are used as the initial visual label L. vis-init ; The audio semantics extraction is as follows; The voiceover A1 is converted into text using a speech recognition model, and the background music style A2 is determined using an audio sentiment analysis model to generate the initial audio tag L. aud-init .

4. The labeling method for an advertising label according to claim 3, characterized in that, In step 3: The semantic association degree between modalities is calculated as follows; Sim calculates the semantic association between text, visual, and audio tags. cross The formula is Where: Sim txt-vis For semantic similarity between text and visual labels; Sim vis-aud For semantic similarity between visual and audio tags; Sim cross The value ranges from 0 to 1, with higher values ​​indicating stronger semantic consistency of multimodal tags. The stage fusion tags are generated as follows; Combining the core objectives of the life cycle stage G stage Sim with cross-modal semantic association cross Generation stage fusion label L stage The formula is L stage ={L|W l ×Sim cross ×G stage ≥0.5}, where W l Assuming weights for the initial labels of a single modality, labels with a comprehensive score ≥ 0.5 are selected as the core labels for this stage.

5. The labeling method for an advertising label according to claim 4, characterized in that, In step 4: The industry-specific dictionary is constructed as follows; For different industries, we collect industry-specific terminology, product characteristic terms, and compliance keywords to build an industry-specific semantic dictionary, Dict. ind It includes three fields: terminology definition, semantic weight, and compliance level; The industry compatibility of the label is calculated as follows; Computational stage fusion label L stage Dict Industry Dictionary ind Adaptability F ind The formula is Where: N match For L stage Number of matches with dictionary terms; W ind The semantic weight of the matching terms; m is the total number of terms in the dictionary; the value ranges from 0 to 1, and the higher the value, the stronger the industry adaptability of the tag; The label calibration procedure is as follows; If F ind ≥0.7, keep L directly. stage If 0.4≤F ind If F < 0.7, supplement the dictionary with terms with high semantic weight; if F ind If the value is less than 0.4, repeat steps 2-3 to optimize the initial tag extraction rules based on the industry dictionary.

6. The labeling method for an advertising label according to claim 5, characterized in that, In step 5: The compliance risk score is calculated as follows; Building a compliance risk thesaurus (Risk) com Calculate the label compliance risk score R com The formula is Where Sim(l,R) com-i W represents the semantic similarity between the tag and the risk word. risk For risk word weights, R com The value range is 0-1; The following privacy risk assessments were conducted. Building a privacy-sensitive word library Risk pri Calculate the privacy risk of the label R pri The formula is Where N pri-match Total represents the number of matches between tags and the privacy thesaurus. l R represents the total number of labels. pri The value range is 0-1; The filtering rules are as follows; If R com >0.3 or R pri If >0, filter the tag directly; if R com ≤0.3 and R pri =0, retain the label and record the compliance score for subsequent performance tracking.

7. The labeling method for an advertising label according to claim 6, characterized in that, In step 6: The performance indicators were collected as follows; Real-time collection of campaign performance data corresponding to tags: Tag exposure (Imp) l Click count l Conversion amount l Calculate the label effect score S e ff, the formula is S e ff = α × CTR l +β×CVR l ; in: Tag click-through rate; Tag conversion rate; α = 0.4, β = 0.6; S eff The value ranges from 0 to 1, representing the actual marketing value of the tag; The label weight iteration is as follows; According to S eff Update tag weight W l -new, the formula is W l -new=W l-old ×(1+γ×S eff ), where W l -old represents the original weights, γ = 0.2 represents the iteration coefficients, and W... l-new The value range is 0-1.5; if S eff If W < 0.2, then reduce the weight W. l -new=W l-old ×0.

8.

8. The labeling method for an advertising label according to claim 7, characterized in that, In step 7: Phase switching detection is as follows; When the advertising campaign reaches the lifecycle stage threshold, or the core objectives of the stage are achieved... Trigger phase switching, update phase identifier S lc ; The tag update process is as follows; After the phase switch, re-execute steps 2-6, based on the new phase's G. stage Industry dictionary and latest performance data are used to generate new stage labels L. s `tag-new` replaces the original tag; at the same time, it retains the historical tag record. h istory={L stage1 ,L stage2 ,L stage3 } is used for retrospective analysis of advertising performance.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory is used to store computer programs; and the processor is used to execute the programs stored in the memory to implement the labeling method of the advertising label as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the labeling method for advertising labels as described in any one of claims 1 to 8.

Images