Internet advertisement pushing method and system

Abstract

The application discloses an internet advertisement pushing method and system, and relates to the technical field of business information systems.The internet advertisement pushing method comprises the following steps: executing the following steps by a discriminator: acquiring creative elements used in advertisement generation and corresponding generated advertisements; generating a target meaning label; generating an unexpected potential meaning label, generating an external environment label, including an external environment label; predicting a delivery risk value of the generated advertisement according to the target meaning label, the unexpected potential meaning label and the external environment label by a preset advertisement risk prediction model; and executing the following steps by a distributor: controlling whether the generated advertisement is delivered according to the predicted delivery risk value and a preset delivery risk threshold value.The method can better push internet advertisements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of business information systems, and in particular to an Internet advertisement pushing method and system. BACKGROUND

[0002] In the field of Internet advertising, Dynamic Creative Optimization (DCO) technology has been widely applied. DCO can dynamically combine and customize the constituent elements (such as scripts, pictures, etc.) in the basic advertisement creative according to real-time hotspots, target audience characteristics, and specific delivery situations, thereby generating a large number of advertisement creative variants.

[0003] Moreover, with the development of Artificial Intelligence Generated Content (AIGC) technology, DCO systems can now use AIGC technology to assist in generating diverse advertisements based on existing material libraries (portraits or artistic materials that have obtained portrait rights or copyright), further enhancing the richness and production efficiency of advertisement creativity. By accurately pushing these advertisement creative variants to different target user groups, the goal is to improve the attractiveness of the advertisement content and key advertisement effect indicators such as user click-through rate and conversion rate.

[0004] However, the existing technology has deficiencies in the content risk assessment process when pushing advertisement creative variants generated using DCO combined with AIGC. Because the generated advertisement creative variants usually need to rely on manual review to identify and exclude potential compliance risks and reputation risks before being pushed. For example, the reviewer needs to rely on experience to judge whether the advertisement content may trigger negative social public opinion, may be improperly associated with negative events, or may touch the sensitive points of specific groups, thereby causing damage to the brand image or product reputation of the advertisement.

[0005] Therefore, the DCO advertisement content risk assessment mechanism in the prior art has the following technical problems: first, the risk assessment process relies on human experience and subjective judgment, lacks objective and unified evaluation standards, and it is difficult to ensure the accuracy and comprehensiveness of Internet advertising risk identification; second, manual review is inefficient and difficult to meet the immediate assessment needs of advertisement creative variants that cater to real-time hotspots, resulting in delayed advertisement delivery, limiting the full play of DCO technology advantages and the pushing efficiency of Internet advertising. SUMMARY

[0006] In order to solve the above technical problems or at least partially solve the above technical problems, the present application provides an Internet advertisement pushing method and system, which can better push Internet advertisements.

[0007] In a first aspect, the present application provides an Internet advertisement pushing method, comprising the following steps:

[0008] The following steps are performed by a discriminator:

[0009] Obtaining the creative elements used in generating the advertisement and the generated advertisement;

[0010] Generating a target meaning label, wherein:

[0011] The target meaning label comprises a creative intention label and a generation compliance label;

[0012] The creative intention label is determined according to the creative elements to describe the creative intention originally expressed by the creative elements;

[0013] The generation compliance label is determined according to the creative elements and the generated advertisement to describe the deviation degree of the generated advertisement from the original advertisement elements;

[0014] Generating an unexpected potential meaning label, wherein:

[0015] Obtaining the similarity between the risk contents in a risk knowledge base and the contents in the generated advertisement, and when the similarity of a risk content in the generated advertisement is greater than a first similarity threshold, setting the risk content and its similarity as the unexpected potential meaning label to describe the risk content unexpectedly implied by the generated advertisement;

[0016] Generating an external environment label, wherein:

[0017] The external environment label is obtained by network data collection to describe the network public opinion environment in which the advertisement group is placed;

[0018] According to the target meaning label, the unexpected potential meaning label and the external environment label, a preset advertisement risk prediction model is used to predict the delivery risk value of the generated advertisement;

[0019] The following steps are performed by a distributor:

[0020] According to the predicted delivery risk value and a preset delivery risk threshold, it is determined whether the generated advertisement is delivered.

[0021] Optionally, the Internet advertisement pushing method further comprises the following steps:

[0022] The discriminator further performs the following steps:

[0023] A second similarity threshold is set, which is less than the first similarity threshold;

[0024] when the similarity of the risk content in the generated advertisement is greater than a second similarity threshold but less than or equal to a first similarity threshold, setting the risk content as a speculative potential meaning label;

[0025] predicting, by a preset advertisement risk prediction model, a speculative delivery risk value of the generated advertisement according to the target meaning label, the speculative potential meaning label, and the external environment label;

[0026] The distributor further performs the following steps:

[0027] detecting, by network data, a discussion heat of the generated advertisement for the speculative potential meaning label, and stopping delivery of the generated advertisement when the discussion heat is greater than a speculative risk heat threshold.

[0028] The speculative risk heat threshold is set according to the speculative delivery risk value, and the greater the speculative delivery risk value, the smaller the speculative risk heat threshold.

[0029] Optionally, the advertisement risk prediction model is constructed by the following steps:

[0030] The advertisement risk prediction model comprises:

[0031] A multi-source labeled information input layer comprising a target meaning label input channel, an unexpected potential meaning label input channel, and an external environment label input channel, for respectively receiving the target meaning label, the unexpected potential meaning label, and the external environment label, and converting the input labels into corresponding label vectors through a word embedding layer;

[0032] A hierarchical risk feature extraction and evaluation layer comprising:

[0033] A direct risk rapid evaluation channel:

[0034] A generation conformity defect evaluation unit composed of a multi-layer perception network, for connecting to the target meaning label input channel to obtain a generation conformity label vector, and outputting a generation conformity defect evaluation feature after processing the generation conformity label vector;

[0035] An unexpected potential meaning risk evaluation unit composed of a multi-layer perception network, for connecting to the unexpected potential meaning label input channel to obtain an unexpected potential meaning label vector, and outputting an unexpected potential meaning risk evaluation feature after processing the unexpected potential meaning label vector;

[0036] A situational interaction risk analysis channel comprising:

[0037] The suitability evaluation unit is composed of a sequentially connected Transformer encoder, a multi-head cross attention mechanism layer and a feedforward fully connected network, and is used to be connected to the target meaning label input channel and the external environment label input channel, to obtain the creative intention label vector and the external environment label vector, and to output a conflict risk feature after processing;

[0038] The multi-channel feature fusion layer is used to be connected to the generation compliance defect evaluation unit, the unexpected potential meaning risk evaluation unit and the suitability evaluation unit respectively, to obtain the generation compliance defect evaluation feature, the unexpected potential meaning risk evaluation feature and the conflict risk feature respectively, and to fuse them into a feature matrix as a multi-channel fusion feature for output;

[0039] The output layer includes a fully connected layer and a Sigmoid activation function, and is used to be connected to the multi-channel feature fusion layer to receive the multi-channel fusion feature, and to process and output an advertisement launching risk value.

[0040] Optionally, the target meaning label is generated, including the following steps:

[0041] The creative elements include advertisement materials and text prompt words;

[0042] Based on the semantic analysis of the text prompt words, a set of creative intention labels for representing the intention of the prompt words are extracted;

[0043] By detecting the similarity between the advertisement materials and the generated advertisement content, a generation compliance label is determined.

[0044] Optionally, the external environment label includes the focus of public opinion, the emotional tendency of the target advertisement launching group, the hot event related to the advertisement element and the emotional tendency corresponding to the hot event.

[0045] Optionally, the external environment label is obtained by the following steps:

[0046] The hot discussion topics of the advertisement launching group in a preset time window and the hot comments under each hot topic are obtained from the advertisement launching media platform;

[0047] The hot discussion topics are subjected to semantic analysis to obtain the focus of public opinion of the target advertisement launching group;

[0048] The hot comments under each hot topic are subjected to sentiment analysis to obtain the emotional tendency of the target advertisement launching group;

[0049] Based on the creative intention label, the hot topics of the public media platform are detected to obtain the hot event related to the advertisement element;

[0050] The sentiment analysis is performed on the hot comments related to the advertising element under the hot spot event, so as to obtain the emotional tendency of the corresponding hot spot event.

[0051] In a second aspect, the present application provides an Internet advertisement pushing system, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the Internet advertisement pushing method according to any one of the first aspect.

[0052] Compared with the prior art, the technical scheme provided by the present application has the following advantages:

[0053] One of the beneficial technical effects and working principles thereof is that:

[0054] In the process of generating advertising variants by using DCO and AIGC technology, three main risks are faced: first, the generated advertising content is inconsistent with the original material, resulting in quality problems such as trademark deformation or face distortion; second, the risk content of accidentally combining negative symbols, sensitive information or easy to trigger associations; third, although the advertising content meets the expectations, it is not coordinated with the current network environment, such as putting light luxury tourism advertising in the public mood low period to cause resentment and attack. The relatively ideal situation is that the advertisement not only has good quality, accurately expresses the creative intention, does not contain risk content, but also can harmoniously coexist with the network public opinion environment and attract the target group.

[0055] In view of these challenges, the present application provides an Internet advertisement pushing method, which discriminates the risk portrait by a discriminator and adjusts the delivery of the advertisement by a distributor accordingly.

[0056] The discriminator first obtains the advertising creative elements and the generation result, and then constructs three types of key labels:

[0057] First, the target meaning label, including the creative intention label determined according to the creative elements and the generation compliance label obtained by comparing the generated advertisement with the advertising material.

[0058] Second, the unexpected potential meaning label, by calculating the similarity between the generated advertising content and the known risk content items in a preset risk knowledge base, when the similarity exceeds a specific first similarity threshold, one or more risk contents and their similarity information matched are determined as one or more unexpected potential meaning labels, so as to reveal the unexpected negative information related to the known risk mode in the advertisement.

[0059] Third, the external environment label, by real-time analysis of network data, a dynamic label describing the network public opinion environment of the target group is formed.

[0060] After these labels are constructed, the preset advertisement risk prediction model comprehensively evaluates the generation quality, potential risks and suitability with the external environment among them, and predicts a quantified advertisement delivery risk value. The distributor then compares the risk value with a delivery risk threshold to determine the delivery strategy.

[0061] Therefore, the present application can systematically cope with various risk situations: by means of the "generation conformity label", the present application can directly identify and cope with the first kind of high risk caused by defects in the AIGC execution level, ensuring that the advertisement content does not deviate from the original intention; moreover, with the help of the "accidental potential meaning label", the present application can actively discover and prevent the second kind of situation where known risk content is accidentally embedded in the advertisement; and by comprehensively evaluating the interaction between the creation intention label and the external environment label, even if the advertisement content itself is well-made and has no obvious accidental risk, the present application can effectively identify and prevent the third kind of advertisement delivery that becomes unsuitable due to conflict with the external environment. On the contrary, in the case of good performance in all dimensions, the model outputs a low risk value, ensuring safe and effective advertisement dissemination.

[0062] Therefore, the present application provides an Internet advertisement pushing method and system, which can better push Internet advertisements.

[0063] The second beneficial technical effect and its working principle are that the advertisements generated by DCO and AIGC technology often contain risk content that is not highly similar to known risk content, but is not completely unrelated, i.e. speculative potential meaning. These speculative potential meanings are not explicitly expressed in the advertisement content, but still have potential risks that are aware of and discussed by the public, and if the first similarity threshold is used, they may be ignored due to not meeting the standard; if the first similarity threshold is lowered, it will also cause a large number of actually acceptable advertisements to be misjudged as high-risk.

[0064] To solve the above problems, the present application introduces a second similarity threshold less than the first similarity threshold at the discriminator level.

[0065] At the discriminator level:

[0066] When the similarity of the advertisement content to a certain item in the risk knowledge base exceeds the second similarity threshold but does not exceed the first similarity threshold, the system marks it as a speculative potential meaning label. Then, the discriminator calls the advertisement risk prediction model to predict a speculative delivery risk value in combination with the target meaning of the advertisement and the external environment label.

[0067] Here, the key role of the speculative launch risk value is that it is not just a risk score, but more importantly, it is used to represent an estimate of the severity of the potential negative consequences that the speculative potential meaning could cause once it is noticed by the public and forms a discussion in the actual launch. In other words, a higher speculative launch risk value means that although the meaning is currently ambiguous or speculative, if it does trigger public opinion, the degree of conflict with the core content of the advertisement, the brand image, or the current social sentiment is expected to be more severe and more likely to cause more widespread or more far-reaching negative effects.

[0068] Therefore, in combination with the processing at the distributor level: for the advertisements containing speculative potential meaning labels and acceptable risk, the system initiates post-monitoring to detect the discussion heat of the speculative risk content in the advertisement launch group in real time. It should be noted that the speculative risk heat threshold used here to determine whether the discussion heat is excessive is dynamically set according to the speculative launch risk value, that is, the more severe the estimated potential consequences, the lower the tolerance of the system to the relevant public opinion, and the smaller the corresponding speculative heat threshold (i.e., making the monitoring system more sensitive). Once the actual discussion heat exceeds this dynamically adjusted threshold, the distributor immediately takes measures such as stopping the launch of the advertisement.

[0069] Therefore, the present application realizes the dynamic correlation of the monitoring sensitivity directly with the severity of the potential consequences of the risk. For risk points that may cause more severe conflicts, the present application can monitor and intervene with higher vigilance; for risk points with lighter potential consequences, relatively lenient standards can be adopted. Thus, the balance between innovative expression and risk control is achieved.

[0070] Therefore, the present application provides an Internet advertisement pushing method and system, which can better push Internet advertisements. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 A demonstration example of the generated Internet advertisement with unexpected negative associations provided for the embodiments of the present application;

[0072] Figure 2 A flowchart of the Internet advertisement pushing method provided for the embodiments of the present application. DETAILED DESCRIPTION

[0073] The technical solutions in the present application will be described below with reference to the accompanying drawings.

[0074] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein; obviously, the embodiments in the description are only some of the embodiments of the present application, not all the embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0075] Referring to Figure 1 As shown in the generated advertisement picture, the hands of the two endorsers are depicted holding the marker in a reverse grip (i.e., with the tail pointing towards the direction of use). The blurred parts are the trademark and the portrait, which are irrelevant to the implementation of the present application.

[0076] In generating the advertisement, the DCO system uses the following reference for the text prompt when generating the picture by AIGC technology: "Two smiling, energetic male endorsers A and young female endorser B are together in a bright and convenient store scene, they are happily chatting, and naturally show the C brand marker D model and E model in their hands to the camera, highlighting the fashion and color diversity of the product." The portrait picture of the corresponding endorser, product picture and manufacturer's trademark are the advertisement elements that the AIGC technology needs to refer to.

[0077] However, the tail of the marker itself has a specific industrial design form (for example, it may be designed to be slightly narrower or have a certain curvature for aesthetics or counterweight), and the trademark part of the product is designed to be printed near the tail, and AIGC may easily output the advertisement of the marker in a reverse grip. Referring to the example, in the finally generated advertisement image, the visual features of the tail of the marker accidentally form a similar shape to the mouthpiece part of the electronic cigarette.

[0078] And this similarity is not derived from the direct instruction of the text prompt in the creative elements of the advertisement, nor is it intended to imitate the electronic cigarette by the design flaw of the marker product. Rather, it is an accidental and unforeseeable implicit content in the original material stage that occurs accidentally after the combination and rendering of object form, holding method and specific observation angle in the complex image generation process.

[0079] And the launch of such an advertisement may lead to inappropriate associations among the target audience (especially parents and educators who are highly sensitive to the health problems of young people), associating stationery products with electronic cigarettes, a type of regulated product, which may negatively impact the brand image of the stationery product or make the advertisement content illegal.

[0080] The internet advertisement pushing method proposed in the present application is aimed at effectively identifying and evaluating various potential risks including such risks.

[0081] Referring to Figure 2 As shown in the generated advertisement picture, the hands of the two endorsers are depicted holding the marker in a reverse grip (i.e., with the tail pointing towards the direction of use). The blurred parts are the trademark and the portrait, which are irrelevant to the implementation of the present application.

[0082] The following steps are performed by the discriminator:

[0083] S1: Obtain the creative elements used in the generation of the advertisement and the corresponding generated advertisement;

[0084] Specifically, the creative elements include advertisement materials (such as preset pictures and brand logo files, etc.) and text prompt words.

[0085] For example, in the embodiments of the present application, the above-mentioned Figure 1 The corresponding advertisement materials include: a portrait photo of male spokesperson A, a portrait photo of young female spokesperson B, a product picture of C brand D model marker, a product picture of C brand E model marker, and a trademark vector graph of C brand manufacturer.

[0086] The text prompt word is "two smiling, sunny, and energetic male spokespersons A and young female spokesperson B are together in a bright convenience store scene, they are happily chatting, and naturally show the C brand marker D model and E model in their hands to the camera, highlighting the fashion and color diversity of the product."

[0087] S2: Generate a target meaning label, wherein:

[0088] The target meaning label includes a creative intention label and a generation compliance label;

[0089] The creative intention label is determined according to the creative elements to describe the creative intention originally expressed by the creative elements;

[0090] Specifically, based on the text prompt word, a semantic analysis is performed to extract a set of creative intention labels for representing the intention of the prompt word.

[0091] In generating the "target meaning label", the determination process of the "creative intention label" specifically includes the following sub-steps:

[0092] Text prompt word preprocessing sub-step: standardization processing is performed on the input text prompt word, such as Chinese word segmentation, part-of-speech tagging, removing general stop words, etc., to form a text sequence suitable for deep semantic analysis.

[0093] Core entity and attribute recognition sub-step: a pre-trained named entity recognition (NER) model and attribute extraction algorithm are used to identify and extract key entity objects (such as spokespersons, product brands, and product models) and their related description attributes (such as spokesperson features "smiling, sunny, and energetic") from the preprocessed text prompt word sequence.

[0094] Creative intention label structured generation sub-step: the information of entities, attributes, actions, scenes and emotions obtained by identification and analysis in the foregoing sub-step is integrated into a set of structured "creative intention labels" according to a predefined label dimension system (for example, subject and feature, core product and characteristic, core action and interaction, scene and atmosphere, composition or visual focus, etc.). For example, the creative intention label generated for the foregoing text prompt is as follows:

[0095] {

[0096] {subject_1_feature: [male, spokesperson A, smiling and sunny, full of vitality]};

[0097] {subject_2_feature: [female, spokesperson B, smiling and sunny, full of vitality]};

[0098] {brand feature: [C brand], product model_feature: [D model, E model]};

[0099] {atmosphere_feature: [happy, sunny, and full of vitality]}.

[0100] }。

[0101] The generation compliance label is determined according to the creative elements and the corresponding generated advertisement to describe the deviation degree of the generated advertisement from the original advertisement elements;

[0102] Specifically, for each original advertisement reference material, the region of interest (refer to the white box region shown in FIG. 1) of the generated advertisement material is detected by a target detection and recognition algorithm (for example, YOLO detection algorithm) in the embodiment of the present application, and the similarity between the corresponding part and the original advertisement material is calculated by SIFT algorithm to determine the generation compliance label. Figure 1

[0103] For example, the generation compliance label generated for the foregoing advertisement material is as follows:

[0104] {

[0105] [male spokesperson A's portrait photo, 0.87];

[0106] [young female spokesperson B's portrait photo, 0.93];

[0107] [product picture of C brand D model marker, 0.83];

[0108] [product picture of C brand E model marker, 0.80];

[0109] [trademark vector graph of C brand manufacturer, 0.99];

[0110] }​

[0111] S3: generating unexpected potential meaning label, wherein:

[0112] Specifically, a risk content knowledge base is pre-constructed and continuously maintained. The knowledge base stores various risk contents that cause adverse effects of advertisements in a structured manner, including prohibited symbols, hate symbols, and inappropriate product images.

[0113] For each risk content, in the embodiment of the present application, a target detection and recognition algorithm (for example, YOLO detection algorithm) is used to detect the region of interest of the risk content in the generated advertisement, and a SIFT algorithm is used to calculate the similarity between the risk content in the risk knowledge base and the content in the generated advertisement. When the similarity of the risk content in the generated advertisement is greater than a first similarity threshold, the risk content and its similarity are set as the unexpected potential meaning label to describe the unexpected risk content implied in the generated advertisement.

[0114] The first similarity threshold is a manually preset value.

[0115] For example, for the above example, the generated unexpected potential meaning label is as follows:

[0116] {

[0117] [Electronic cigarette, 0.74];

[0118] }.

[0119] S4: generating an external environment label, wherein:

[0120] The external environment label is obtained through network data collection to describe the network public opinion environment in which the target advertising group is located.

[0121] Specifically, the external environment label includes the focus of public opinion, the emotional tendency, the hot events related to the advertising elements, and the emotional tendency corresponding to the hot events of the target advertising group.

[0122] Specifically, the external environment label is obtained through the following steps:

[0123] Obtain the hot discussion topics of the advertising group in a preset time window and the hot comments under each hot topic from the advertising media platform.

[0124] Perform semantic analysis on the hot discussion topics to obtain the current hot keywords as the focus of public opinion.

[0125] Perform sentiment analysis on the hot comments under each hot topic to obtain the emotional tendency of the target advertising group.

[0126] Detecting hot topics of a public media platform based on the creative intention label, thereby obtaining an advertisement element related hot keyword as a hot event;

[0127] Performing sentiment analysis on hot comments under the advertisement element related hot event, thereby obtaining an emotional tendency of the corresponding hot event.

[0128] Specifically, from one or more publicly accessible network data sources (e.g., mainstream social media platforms such as Weibo, WeChat public account, etc.), obtain hot discussion topics and representative user comments or content under each hot topic within a preset time window related to the target audience of the advertisement.

[0129] Specifically:

[0130] Set a preset time window (such as the past 24 hours, the past week), continuously collect hot topics and / or articles and related comments data related to the target audience of the advertisement.

[0131] Use natural language processing techniques (such as topic modeling algorithm LDA, text clustering, keyword extraction) to perform semantic analysis and aggregation on hot discussion topics, to identify and determine the main "public opinion focus" of the target audience within the current time window and its representative keywords.

[0132] Further perform sentiment analysis (e.g., deep learning based sentiment classification model or sentiment dictionary method) on representative user comments under these hot discussion topics, and statistically determine the "emotional tendency" of the target audience of the advertisement to these focus and its intensity.

[0133] Based on the creative intention label of the advertisement, use keyword matching technology to actively discover associated discussion topics and their representative keywords, calculate user interaction indicators (such as comments, likes, forwards, etc.) of the discussion topics, for example, add these interaction indicators according to a certain weight ratio to obtain the heat index of each discussion topic, and the discussion topics with a heat index greater than a preset heat threshold are considered as hot events.

[0134] Specifically, the preset heat threshold is a manually set value.

[0135] For each identified "advertisement element related hot event", collect and analyze user comments directly associated with it, and obtain the "emotional tendency" of the public to each hot event and its intensity through sentiment analysis technology.

[0136] Integrate the public opinion focus and "emotional tendency, as well as the advertisement element related hot event and its emotional tendency into an external environment label in the form of a pre-defined structure or dynamically generated key-value pair.

[0137] For example, an example of a complete external environment tag is:

[0138] {

[0139] Target group focus: [Keyword A, Keyword B];

[0140] Target group sentiment: [Positive: 0.6, Negative: 0.2, Neutral: 0.2];

[0141] Associated hot event: [Keyword C, Keyword D];

[0142] Associated hot event sentiment: [Positive: 0.3, Negative: 0.6, Neutral: 0.1]

[0143] }.

[0144] S5: predicting the delivery risk value of the generated advertisement according to the target meaning tag, the unexpected potential meaning tag and the external environment tag through a preset advertisement risk prediction model;

[0145] Specifically, the advertisement risk prediction model is constructed by the following steps:

[0146] The advertisement risk prediction model includes:

[0147] A multi-source tagged information input layer including a target meaning tag input channel, an unexpected potential meaning tag input channel and an external environment tag input channel, for receiving the target meaning tag, the unexpected potential meaning tag and the external environment tag respectively, and converting the inputted each tag into a corresponding tag vector through a word embedding layer;

[0148] A hierarchical risk feature extraction and evaluation layer, including:

[0149] A direct risk rapid evaluation channel:

[0150] A generation conformity defect evaluation unit composed of a multi-layer perception network, for connecting to the target meaning tag input channel to obtain a generation conformity tag vector, and outputting a generation conformity defect evaluation feature after processing the generation conformity tag vector;

[0151] An unexpected potential meaning risk evaluation unit composed of a multi-layer perception network, for connecting to the unexpected potential meaning tag input channel to obtain an unexpected potential meaning tag vector, and outputting an unexpected potential meaning risk evaluation feature after processing the unexpected potential meaning tag vector;

[0152] A situational interaction risk analysis channel, including:

[0153] The suitability evaluation unit is composed of a sequentially connected Transformer encoder, a multi-head cross attention mechanism layer, and a feedforward fully connected network, and is used to be connected to the target meaning label input channel and the external environment label input channel to obtain the creative intention label vector and the external environment label vector and output a conflict risk feature after processing;

[0154] The multi-channel feature fusion layer is used to be connected to the generation conformity defect evaluation unit, the unexpected potential meaning risk evaluation unit, and the suitability evaluation unit to obtain generation conformity defect evaluation features, unexpected potential meaning risk evaluation features, and conflict risk features, respectively, and fuse them into a feature matrix as a multi-channel fusion feature for output;

[0155] The output layer includes a fully connected layer and a Sigmoid activation function and is used to be connected to the multi-channel feature fusion layer to receive the multi-channel fusion feature, process it, and output an advertisement launching risk value.

[0156] Specifically, in the embodiment of the present application, the advertisement risk prediction model is obtained by the following steps:

[0157] A plurality of application AIGC technology generated advertisement sample input features are collected as inputs of the advertisement risk prediction model, and their corresponding advertisement launching risk values are used as labels for training the advertisement risk prediction model, so as to train the advertisement risk prediction model in a supervised learning manner:

[0158] For each advertisement sample in the training set, its input features are composed of three groups of labels (target meaning labels, unexpected potential meaning labels, and external environment labels) corresponding to the advertisement sample obtained by the same method as in the above embodiment.

[0159] The advertisement launching risk value is a standardized continuous value between 0.0 and 1.0, and the higher the value, the greater the risk. The advertisement launching risk value corresponding to each advertisement sample in the training set is given by a labeling expert with professional knowledge of advertisement review according to the risk score given by the advertisement sample for each advertisement sample.

[0160] The following steps are performed by the distributor:

[0161] S6: According to the predicted launching risk value and the preset launching risk threshold, it is controlled whether the generated advertisement is launched.

[0162] Specifically, when the launching risk value is greater than the preset launching risk threshold, the launching of the generated advertisement is rejected.

[0163] Specifically, the launching risk threshold is a manually set value.

[0164] In another embodiment, the Internet advertisement pushing method further comprises the following steps:

[0165] The discriminator further performs the following steps:

[0166] A second similarity threshold is set, which is less than the first similarity threshold;

[0167] When the similarity of the risk content in the generated advertisement is greater than the second similarity threshold but less than or equal to the first similarity threshold, the risk content is set as a speculative potential meaning label.

[0168] The second similarity threshold is a manually set value.

[0169] By a preset advertisement risk prediction model, the speculative delivery risk value of the generated advertisement is predicted according to the target meaning label, the speculative potential meaning label, and the external environment label.

[0170] Specifically, when there are multiple speculative potential meaning labels, each speculative potential meaning label is respectively combined with the target meaning label and the external environment label, and then input into the advertisement risk prediction model to predict the speculative delivery risk value of the generated advertisement containing each speculative potential meaning label.

[0171] The distributor further performs the following steps:

[0172] The discussion heat of the generated advertisement for the speculative potential meaning label is detected through network data, and when the discussion heat is greater than a speculative risk heat threshold, the delivery of the generated advertisement is stopped.

[0173] Using a keyword matching technology, a discussion topic associated with the speculative potential meaning label is actively discovered, and a heat index of each discussion topic associated with the speculative potential meaning label is obtained by calculating user interaction indicators (such as comments, likes, and forwards) of the discussion topic, for example, by adding these interaction indicators in a certain weight ratio.

[0174] The speculative risk heat threshold is set according to the speculative delivery risk value, and the greater the speculative delivery risk value, the smaller the speculative risk heat threshold.

[0175] Specifically, the speculative risk heat threshold is determined by the following formula and the speculative delivery risk value:

[0176]

[0177] wherein, is the speculative risk heat threshold, a lowest risk heat threshold set by human, a highest risk heat threshold set by human, a speculative delivery risk value, an empirical parameter set by human.

[0178] One of the beneficial technical effects of the embodiments of the present application and the working principle is that:

[0179] In the process of generating advertising variants using DCO and AIGC technology, there are three main risks: first, the generated advertising content is inconsistent with the original material, resulting in quality problems such as trademark deformation or face distortion; second, accidentally combining negative symbols, sensitive information or risk content that can easily trigger associations; third, the advertising content may be consistent with expectations but not in harmony with the current network environment, such as launching a light luxury travel advertisement during a period of public low morale, which may cause resentment and attacks. The relatively ideal situation is that the advertisement not only has excellent quality, accurately expresses the creative intent, and does not contain risk content, but also can coexist harmoniously with the network public opinion environment and attract the target group.

[0180] To address these challenges, the present application provides an Internet advertising delivery method, which uses a discriminator to perform risk profiling and a distributor to regulate the delivery of advertisements accordingly.

[0181] The discriminator first obtains the advertising creative elements and the generation result, and then constructs three types of key labels:

[0182] First, the target meaning label, which includes the creative intent label determined according to the creative elements and the generation compliance label obtained by comparing the generated advertisement with the advertising material.

[0183] Second, the unexpected potential meaning label, which is determined by calculating the similarity between the generated advertising content and the known risk content items in a pre-set risk knowledge base. When the similarity exceeds a specific first similarity threshold, the matched one or more risk contents and their similarity information are collectively determined as one or more unexpected potential meaning labels, revealing the potential unexpected negative information related to known risk patterns in the advertisement.

[0184] Third, the external environment label, which is formed by real-time analysis of network data to describe the dynamic label of the network public opinion environment in which the target group is located.

[0185] After these labels are constructed, the pre-set advertising risk prediction model will comprehensively evaluate their generation quality, potential risks and suitability with the external environment to predict a quantitative advertising delivery risk value. The distributor will then compare this risk value with the delivery risk threshold to determine the delivery strategy.

[0186] Therefore, the application can systematically deal with various risk situations: by generating the conformity label, the application can directly identify and deal with the first high risk caused by defects in the AIGC execution level, ensuring that the advertising content does not deviate from the original intention; moreover, with the help of the unexpected potential meaning label, the application can actively discover and prevent the second situation of accidentally embedding known risk content in the advertisement; and by comprehensively evaluating the interaction of the "creation intention label" and the "external environment label", even if the advertising content itself is well-made and has no obvious unexpected risks, the application can effectively identify and prevent the third situation of becoming inappropriate due to conflict with the external environment. On the contrary, in the case of good performance in all dimensions, the model outputs a low risk value, ensuring safe and effective advertising dissemination.

[0187] Therefore, the application provides an Internet advertisement pushing method and system, which can better push Internet advertisements.

[0188] The second beneficial technical effect of the embodiment of the application and its working principle is that the advertisements generated by the DCO and AIGC technologies often have risk content that is not highly similar to known risk content, but is not completely unrelated, that is, the inferred potential meaning. These inferred potential meanings are not explicitly expressed in the advertising content, but still have potential risks that are aware of and discussed by the public, and if the first similarity threshold is used for judgment, they may be ignored due to not meeting the standard; if the first similarity threshold is lowered, a large number of actually acceptable advertisements may be misjudged as high-risk.

[0189] To solve the above problems, the application introduces a second similarity threshold less than the first similarity threshold at the discriminator level.

[0190] At the discriminator level:

[0191] When the similarity of the advertising content to a certain item in the risk knowledge base exceeds the second similarity threshold but does not exceed the first similarity threshold, the system marks it as an inferred potential meaning label. Then, the discriminator calls the advertising risk prediction model to predict the inferred risk value of the advertisement in combination with the target meaning and external environment label of the advertisement.

[0192] Here, the key role of the inferred risk value is not only a risk score, but more importantly, it is used to represent the inferred potential meaning, and once it is noticed by the public and forms a discussion in actual delivery, it is a prediction of the severity of the potential negative consequences it may cause. In other words, a higher "inferred risk value" means that although the meaning is currently ambiguous or inferred, if it really triggers public opinion, the degree of conflict with the core content of the advertisement, the brand image, or the current social sentiment is expected to be more serious, and it is more likely to cause more widespread or more far-reaching negative effects.

[0193] Therefore, in combination with the processing at the distributor level: for the ads containing the speculative potential implication label and the acceptable risk, the system initiates the post-monitoring, which detects the discussion heat of the speculative risk content in the ad delivery group in real time. It should be noted that the speculative risk heat threshold used to determine whether the discussion heat is excessive is dynamically set according to the speculative delivery risk value, that is, the more serious the estimated potential consequences, the lower the tolerance of the system to the relevant public opinion, and the smaller the corresponding speculative heat threshold (that is, to make the monitoring system more sensitive). Once the actual discussion heat exceeds the dynamically adjusted threshold, the distributor immediately takes measures such as stopping the ad delivery.

[0194] Therefore, the present application realizes the dynamic correlation of the monitoring sensitivity and the severity of the potential consequences of the risk. For the risk points that may cause more intense conflicts, the present application can monitor and intervene with higher vigilance; for the risk points with lighter potential consequences, relatively lenient standards can be adopted. Thus, the balance between innovative expression and risk control is achieved.

[0195] Therefore, the present application provides an Internet ad pushing method and system, which can better push Internet ads.

[0196] The embodiment of the present application also provides an Internet ad pushing system, which comprises a processor and a memory, and the memory stores at least one instruction, at least one program, a code set or an instruction set. The at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to realize the Internet ad pushing method as described in any one of the above embodiments.

[0197] It should be noted that the embodiments described in the present application have been subjected to corresponding compliance measures before actual deployment and application, to ensure compliance with relevant laws and regulations, social morality and public interest requirements. Especially when the technical solutions of the present application involve algorithm features, business rules and methods, such as data collection, label management, rule setting and / or recommendation decision, etc. specific aspects, necessary compliance measures have been fully taken, including but not limited to: establishing effective data use agreements with the subjects involved in the data to fulfill the privacy protection obligations; adjusting the model training data as necessary to avoid or eliminate possible bias; and strictly complying with relevant national laws and regulations when executing the method steps, to maintain public interest and market order. However, since such compliance measures belong to the category that can be selected and implemented by those skilled in the art based on ordinary technical knowledge and conventional experimental means, they do not constitute the necessary technical features for implementing the technical solutions of the present application.

[0198] It should be noted that, in the present document, the terms such as "first" and "second" and the like are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual such relationship or order between such entities or operations. In addition, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the stated element. Also, in the description of the embodiments of the present application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in the present document is only a description of the associated relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. Also, in the description of the embodiments of the present application, "a plurality of" means two or more than two.

[0199] The above description is only a specific implementation of the present application, which enables those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments described herein, but will conform to the widest range consistent with the principles and novel features disclosed herein.

Claims

1. An Internet advertisement pushing method, characterized by, The internet advertisement pushing method comprises the following steps: The following steps are performed by the discriminator: Obtaining the creative elements used in the generation of the advertisement and the corresponding generated advertisement; Generating a target meaning label, wherein: The target meaning label comprises a creative intention label and a generation compliance label; The creative intention label is determined according to the creative elements to describe the creative intention originally expressed by the creative elements; The generation compliance label is determined according to the creative elements and the corresponding generated advertisement to describe the degree of deviation of the generated advertisement from the original advertisement elements; Generating an unexpected potential meaning label, wherein: Obtaining the similarity between the risk content in the risk knowledge base and the content in the generated advertisement, and when the similarity of the risk content in the generated advertisement is greater than a first similarity threshold, setting the risk content and its similarity as the unexpected potential meaning label to describe the risk content unexpectedly implied by the generated advertisement; Generating an external environment label, wherein: The external environment label is obtained through network data collection to describe the network public opinion environment in which the advertisement group is placed; Predicting the delivery risk value of the generated advertisement according to the target meaning label, the unexpected potential meaning label and the external environment label through a preset advertisement risk prediction model; The following steps are performed by the distributor: According to the predicted delivery risk value and a preset delivery risk threshold, controlling whether to deliver the generated advertisement.

2. The Internet advertisement pushing method of claim 1, wherein, The internet advertisement pushing method further comprises the following steps: The discriminator further performs the following steps: Setting a second similarity threshold, which is less than the first similarity threshold; When the similarity of the risk content in the generated advertisement is greater than the second similarity threshold but less than or equal to the first similarity threshold, setting the risk content as a speculative potential meaning label; Predicting the speculative delivery risk value of the generated advertisement according to the target meaning label, the speculative potential meaning label and the external environment label through a preset advertisement risk prediction model; The distributor further performs the following steps: Stopping the delivery of the generated advertisement when the discussion heat of the speculative potential meaning label by the delivered advertisement group is greater than a speculative risk heat threshold through network data detection; Wherein, the speculative risk heat threshold is set according to the speculative delivery risk value, and the greater the speculative delivery risk value, the smaller the speculative risk heat threshold.

3. The Internet advertisement pushing method of claim 1, wherein, The advertisement risk prediction model is constructed by the following steps: The advertisement risk prediction model comprises: A multi-source labeled information input layer comprising a target meaning label input channel, an unexpected potential meaning label input channel and an external environment label input channel for respectively receiving the target meaning label, the unexpected potential meaning label and the external environment label, and converting the input labels into corresponding label vectors through a word embedding layer; A hierarchical risk feature extraction and evaluation layer comprising: A direct risk rapid evaluation channel: The generation compliance defect evaluation unit is composed of a multi-layer perception network and is connected to the target meaning label input channel to obtain a generation compliance label vector. After processing the generation compliance label vector, a generation compliance defect evaluation feature is output. The unexpected potential meaning risk evaluation unit is composed of a multi-layer perception network and is connected to the unexpected potential meaning label input channel to obtain an unexpected potential meaning label vector. After processing the unexpected potential meaning label vector, an unexpected potential meaning risk evaluation feature is output. The situational interaction risk analysis channel includes: The suitability evaluation unit is composed of a sequentially connected Transformer encoder, a multi-head cross-attention mechanism layer, and a feedforward fully connected network. It is connected to the target meaning label input channel and the external environment label input channel to obtain the creative intention label vector and the external environment label vector. After processing, a conflict risk feature is output. The multi-channel feature fusion layer is connected to the generation compliance defect evaluation unit, the unexpected potential meaning risk evaluation unit, and the suitability evaluation unit to obtain the generation compliance defect evaluation feature, the unexpected potential meaning risk evaluation feature, and the conflict risk feature, respectively. These features are fused into a feature matrix as a multi-channel fusion feature for output. The output layer includes a fully connected layer and a Sigmoid activation function. It is connected to the multi-channel feature fusion layer to receive the multi-channel fusion feature, process it, and output an advertisement delivery risk value.

4. The Internet advertisement pushing method of claim 1, wherein, The target meaning label includes the following steps: The creative elements include advertisement materials and text prompt words. Based on the semantic analysis of the text prompt words, a set of creative intention labels representing the intention of the prompt words is extracted. The generation compliance label is determined by detecting the similarity between the advertisement materials and the generated advertisement content.

5. The Internet advertisement pushing method of claim 4, wherein, The external environment label includes the focus of public opinion, emotional tendency, advertisement element related hot events, and emotional tendency corresponding to the hot events of the target advertisement delivery group.

6. The Internet advertisement pushing method of claim 5, wherein, The external environment label is obtained by the following steps: From the advertisement delivery media platform, the hot discussion topics and hot comments under each hot topic of the advertisement delivery group within a preset time window are obtained. The semantic analysis of the hot discussion topics obtains the focus of public opinion of the advertisement target delivery group. The sentiment analysis of the hot comments under each hot topic obtains the emotional tendency of the advertisement target delivery group. Based on the creative intention label, the hot topics of the public media platform are detected to obtain the advertisement element related hot events. The sentiment analysis of the hot comments under the advertisement element related hot events obtains the emotional tendency corresponding to the hot events.

7. An Internet advertisement pushing system characterized by comprising: The internet advertisement pushing system includes a processor and a memory. The memory stores at least one instruction, at least one program, a code set, or an instruction set. The at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the internet advertisement pushing method as claimed in any one of claims 1-6.

Images