Advertisement classification method, system, equipment and medium

Abstract

The invention discloses an advertisement classification method, system, device and medium, and the method comprises the steps: firstly obtaining the advertisement content of a platform side and the classification label content of an advertiser side, and then processing the advertisement content to obtain an advertisement description text and a picture understanding text, a video understanding text or a video subtitle text; splicing the advertisement description text with a picture understanding text, a video understanding text or a video caption text to obtain an advertisement input text, constructing a multi-level classification system according to classification label contents, extracting classification feature vectors of all levels through a text feature extraction model based on the multi-level classification system, and obtaining a classification feature vector of all levels; and finally, through a pre-trained multi-level context linkage learning model, providing context information for subsequent classification according to a preorder classification result in an iteration mode, determining classification affiliation of the advertisement input text in each level of classification feature vector level by level, and generating an advertisement classification result. According to the invention, the efficiency and accuracy of advertisement classification can be improved.

Description

Technical Field

[0001] This invention relates to the field of text classification and processing technology, and in particular to an advertising classification method, system, device and medium. Background Technology

[0002] With the advent of the digital media era, advertising channels have rapidly expanded from traditional television media to various online platforms. Currently, many platforms rely on advertisers to set their own ad categories or for manual review. This method is not only inefficient but also highly susceptible to errors due to subjectivity or cognitive biases. Incorrect categorization directly affects the attribution analysis of advertising effectiveness and leads to misjudgments in subsequent advertising strategies, causing adjustments to these strategies to deviate from actual needs.

[0003] To address these shortcomings, machine learning-based advertising classification solutions have emerged in the industry. For example, Chinese patent CN115186117A discloses an advertising classification method that crawls advertising data from open platforms, parses the text information, performs zero-shot unsupervised pre-classification, then combines manual verification to create a training dataset, and finally uses an accelerated pre-trained model to train a three-level classification model. While this technology improves classification efficiency to some extent, it still cannot meet the complex needs of the current advertising industry. Its core relies on predefined advertising data types and classification systems. Model training and inference revolve around fixed categories, resulting in extremely poor adaptability in scenarios where advertising categories are dynamically adjusted according to market demand or advertising objectives are rapidly iterated. It struggles to complete the classification task of adding or iterating new categories. Furthermore, this technology isolates the classification processes of different levels and categories, ignoring the close logical connections between multi-level labels in advertising classification. This leads to the loss of key contextual information during the classification process, easily resulting in logically contradictory classification results and affecting the overall classification performance. In addition, the pre-trained model it uses is relatively simple and lacks the ability to distinguish between subcategories. In scenarios where high accuracy is required for single-level classification and overall consistency across multi-level classifications are needed, its accuracy performance cannot meet the needs of practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide an advertising classification method, system, device, and medium to improve the efficiency and accuracy of advertising classification.

[0005] To achieve the above objectives, the present invention provides an advertising classification method, comprising: Obtain advertising content from the platform and category tag content from the advertiser; The advertising content is processed to obtain advertising description text and image understanding text, video understanding text or video subtitle text. The advertising description text is then concatenated with the image understanding text, video understanding text or video subtitle text to obtain the advertising input text. A multi-level classification system is constructed based on the content of the classification labels, and the classification feature vectors of each level are extracted based on the multi-level classification system using a text feature extraction model. By using a pre-trained multi-level contextual learning model, contextual information is provided for subsequent classifications based on the previous classification results in an iterative manner. The classification of the input text for the advertisement is determined level by level in the classification feature vectors of each level, thereby generating the advertisement classification result.

[0006] Preferably, the step of using a pre-trained multi-level contextual learning model to iteratively provide contextual information for subsequent classifications based on previous classification results, and to progressively determine the classification affiliation of the advertisement input text in the classification feature vectors at each level to generate an advertisement classification result, includes: The feature vector of the advertising input text is extracted by the text feature extraction module and used as the first-level input feature vector of the multi-level context linkage learning model. For each classification level, the input feature vector of the current level is matched with the feature vectors of each level to select the target classification for the current level; based on the input feature vector of the current level and the feature vector of the target classification, the input feature vector of the next level is generated. Continue until the last level of classification, and use the classification result of the last level as the advertising classification result.

[0007] Preferably, the multi-level context-linked learning model is obtained through the following steps: Obtain advertising data, select a portion of training data from the advertising data to train a multi-level context linkage learning network, and obtain a primary multi-level context linkage learning model; The remaining advertising data is iteratively and hierarchically optimized using the aforementioned primary multi-level contextual learning model to construct an advertising training dataset. The advertising training dataset is used as the input to the primary multi-level context linkage learning model. The primary multi-level context linkage learning model is trained by hierarchical classification using a joint loss function to obtain a trained multi-level context linkage learning model.

[0008] Preferably, the joint loss function includes a single-level loss function for optimizing the classification accuracy at each level, and a global loss function for constraining the consistency of classification results at all levels.

[0009] Preferably, the step of using a joint loss function to perform hierarchical classification training on the primary multi-level context linkage learning model to obtain a trained multi-level context linkage learning model includes: The features output from the previous level classification are mixed with the current level's true label features according to preset weights to construct the next level training input features for the next level classification training. A single-level loss function is used to optimize the classification accuracy at each level, and a global loss function is used to constrain the consistency of classification results at all levels until all levels of classification training are completed, resulting in a trained multi-level context linkage learning model.

[0010] Preferably, the step of processing the advertisement content to obtain advertisement description text, image understanding text, or video understanding text, and concatenating the advertisement description text with the image understanding text, video understanding text, or video subtitle text to obtain the advertisement input text, includes: The advertisement content is segmented and cleaned to obtain the advertisement description text; When the advertising content contains images or videos, the keyframes of the images or videos are understood through a visual language model, and the understanding results are segmented and cleaned to obtain image-understood text or video-understood text; the subtitles of the video image frames are extracted through a text recognition model, and the extracted subtitles are segmented and cleaned to obtain video subtitle text. The advertisement description text is concatenated with the image understanding text, video understanding text, or video subtitle text to obtain the advertisement input text.

[0011] Preferably, the visual language model is one of a contrastive language-image pre-trained model, an image understanding model, or a guided language-image pre-trained model; the character recognition model is an optical character recognition model.

[0012] To achieve the above objectives, the present invention also provides an advertising classification system, comprising: The data acquisition module is used to acquire advertising content from the platform side and category tag content from the advertiser side. The data processing module is used to process the advertising content to obtain advertising description text and image understanding text, video understanding text or video subtitle text, and to concatenate the advertising description text with the image understanding text, video understanding text or video subtitle text to obtain the advertising input text; The classification feature extraction module is used to construct a multi-level classification system based on the content of the classification labels, and extract classification feature vectors at each level based on the multi-level classification system using a text feature extraction model. The advertising classification module is used to provide contextual information for subsequent classifications based on the previous classification results through a pre-trained multi-level contextual linkage learning model. It determines the classification of the advertising input text in the classification feature vectors of each level step by step, and generates advertising classification results.

[0013] To achieve the above objectives, the present invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the advertising classification method as described above.

[0014] To achieve the above objectives, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the advertising classification method as described above.

[0015] Compared with existing technologies, the advertising classification method, system, device, and medium provided by this invention extracts dynamically associative classification feature vectors and combines them with hierarchical inference of a multi-level context-linked learning model. This allows for efficient handling of dynamically changing advertising categories without relying on predefined classification categories. By introducing a weighted feature fusion mechanism between levels during model training and inference, a tight logical association between multi-level classification labels is established, avoiding the loss of key contextual information and logical contradictions. Through a combination of single-level precise learning and global learning, the model's classification capabilities are comprehensively improved, making it more robust in complex business scenarios with high requirements for classification accuracy and consistency. The embodiments of this invention achieve high efficiency and accuracy in advertising classification, effectively solving the problem of classification errors. They provide reliable data support for advertising performance attribution analysis, ensuring that subsequent adjustments to advertising placement strategies align with actual needs, thereby improving the robustness of advertising placement. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an advertising classification method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the training process architecture of a multi-level context-linked learning model provided in an embodiment of the present invention; Figure 3 This is a structural block diagram of an advertising classification system provided in an embodiment of the present invention; Figure 4 This is a structural block diagram of a terminal device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 , Figure 1 This is a flowchart of an advertising classification method provided by an embodiment of the present invention, the advertising classification method including steps S1 to S4: Step S1: Obtain the advertising content from the platform side and the category tag content from the advertiser side; For example, an ad data source set is constructed by pulling the URLs of the ads to be placed from the media platform. Then based on the ad data source set The webpage content is parsed from each URL to obtain the content corresponding to each ad, which includes the information time, copy, and data source (image or video); the corresponding brand category tag content is obtained from the advertiser's side.

[0020] Step S2: Process the advertisement content to obtain advertisement description text and image understanding text, video understanding text or video subtitle text. Concatenate the advertisement description text with the image understanding text, video understanding text or video subtitle text to obtain the advertisement input text. In an optional embodiment, step S2 includes steps S201 to S203: Step S201: Perform word segmentation and content cleaning on the advertising content to obtain the advertising description text; For example, the advertising content obtained in step S1 is segmented into words, and then the segmentation results are cleaned and filtered to remove information related to non-core advertising content, such as weather, environment, interjections, and titles, to obtain the advertising description text. .

[0021] Step S202: When the advertising content contains images or videos, the keyframes of the images or videos are understood through a visual language model, and the understanding results are segmented and cleaned to obtain image understanding text or video understanding text; the subtitles of the video image frames are extracted through a text recognition model, and the extracted subtitles are segmented and cleaned to obtain video subtitle text. For example, this can be achieved by analyzing the set of advertising data sources in step S1. Process it; If the data source is an image, a visual language model is needed to understand the image. The understanding result is then segmented into words, and non-core advertising content such as weather, environment, interjections, and titles is removed from the segmented results to obtain the image-understood text. ; If the data source is video, the key image frames of the video are first obtained by calculating the differences and similarities between consecutive video frames. Then, the images are understood frame by frame using a text recognition model. The segmentation results are then processed to remove non-advertising core content-related information, such as weather, environment, interjections, and titles, to obtain the video-understood text. The text is then processed using a text recognition model to extract subtitles from video image frames. The obtained subtitles are segmented into words, and non-advertising content is removed to obtain the video subtitle text. .

[0022] Preferably, the visual language model is one of the following: Contrastive Language-Image Pre-training (CLIP), image understanding model, or Bootstrapped Language-Image Pre-training (BLIP).

[0023] Preferably, the character recognition model is an optical character recognition (ORC) model.

[0024] Step S203: Concatenate the advertisement description text with the image understanding text, video understanding text, or video subtitle text to obtain the advertisement input text.

[0025] For example, the advertisement description text Understanding text with images Or video to understand text or video subtitle text By concatenating the text, we obtain the advertisement input text. .

[0026] Step S3: Construct a multi-level classification system based on the classification tag content, and extract classification feature vectors at each level based on the multi-level classification system using a text feature extraction model; For example, the category tags provided by the advertiser are divided into first-level, second-level, ..., etc.; for the automotive scenario: manufacturer, brand, series, level, price, number of seats, energy, and driving range, a total of 8 levels. Then, the divided and registered content is categorized and associated, such as: Car A, Little A, SU7 2024 model, sedan, starting from 200,000, 5-seater, new energy, starting from 700km; AITO Wenjie, Wenjie, M7 2024 model, SUV, starting from 250,000, 5-seater, new energy, starting from 230km; Car B, Little B, Cena 2026 model, starting from 300,000, 7-seater, gasoline car, 1200km, etc. Finally, all levels of classification are feature-encoded using a text feature extraction model to obtain corresponding feature vectors. In practice, these feature vectors can be stored in a feature vector library to obtain a classification vector library. .

[0027] Step S4: Using a pre-trained multi-level contextual linkage learning model, contextual information is provided for subsequent classifications based on the previous classification results in an iterative manner. The classification of the input text for the advertisement is determined level by level in the classification feature vectors of each level, and the advertisement classification result is generated.

[0028] In an optional embodiment, step S4 includes steps S401 to S403: Step S401: Extract the feature vector of the advertisement input text through the text feature extraction module, and use it as the first-level input feature vector of the multi-level context linkage learning model; Step S402: For each classification level, match the input feature vector of the current level with the feature vectors of each level to select the target classification of the current level; based on the input feature vector of the current level and the feature vector of the target classification, generate the input feature vector of the next level. Step S403: Continue until the last level of classification, and use the classification result output at the last level as the advertising classification result.

[0029] For example, the advertisement is entered as text. Feature extraction is performed using a text feature extraction model to obtain the feature vector of the advertisement input text. Then the advertisement is input into the text feature vector. With classification vector library The cosine similarity (cosine) is calculated for each feature vector of the same level. , The algorithm takes the target category with the highest similarity as the current feature target, calculates whether the feature target is consistent with the encoder output target of the multi-level context linkage learning model, and calculates the corresponding weight W_G_1. Then, it takes the vector of the current feature target, multiplies it by the weight W_G_1, takes the first-level classification feature vector, multiplies it by the weight W_1, adds them together, and then uses them as the feature vector to pass into the next level of classification. This process is repeated 7 times until the 8th level of target classification. The 8th level classification directly uses the result of the encoder output as the final advertising classification result.

[0030] Preferably, all category results can be output and displayed according to different modes. For example, the category results of each advertisement can be progressively filled into text or table-type output documents such as txt, word, and xlsx, and the output file can be returned to the user.

[0031] In one optional embodiment, the multi-level context-linked learning model is obtained through the following steps: Obtain advertising data, select a portion of training data from the advertising data to train a multi-level context linkage learning network, and obtain a primary multi-level context linkage learning model; The remaining advertising data is iteratively and hierarchically optimized using the aforementioned primary multi-level contextual learning model to construct an advertising training dataset. The advertising training dataset is used as the input to the primary multi-level context linkage learning model. The primary multi-level context linkage learning model is trained by hierarchical classification using a joint loss function to obtain a trained multi-level context linkage learning model.

[0032] Specifically, the joint loss function includes a single-level loss function for optimizing the classification accuracy at each level, and a global loss function for constraining the consistency of classification results across all levels.

[0033] Furthermore, the step of using a joint loss function to perform hierarchical classification training on the primary multi-level context linkage learning model to obtain a trained multi-level context linkage learning model includes: The features output from the previous level classification are mixed with the current level's true label features according to preset weights to construct the next level training input features for the next level classification training. A single-level loss function is used to optimize the classification accuracy at each level, and a global loss function is used to constrain the consistency of classification results at all levels until all levels of classification training are completed, resulting in a trained multi-level context linkage learning model.

[0034] For example, firstly, relevant ad links are crawled from major media platforms, and then the content of the current ad is parsed link by link, including the title, copy, image or video, to obtain ad data and complete the collection of training data; The collected advertising data is preprocessed to obtain ; First, through manual labor A small batch of training data, approximately 1000 records, was cleaned and then fed into a multi-level contextual learning network (MLN) to obtain a preliminary MLF model. This preliminary MLF model was then used to perform preliminary classification on the remaining data, yielding approximately 10,000 records. These 10,000 records were then manually cleaned to produce high-quality data. Finally, through approximately six iterations using a combination of model and manual cleaning, only 600,000 training records were obtained, forming the advertising training dataset. This dataset was then split into training, validation, and test sets in a 9:1:1 ratio.

[0035] The training set is then fed into a primary multi-level contextual learning model for training, resulting in a final classification depth of 8 different levels of classification, such as automotive scenarios, manufacturer, brand, series, class, price, seats, energy, and driving range. It should be noted that the network in this embodiment of the invention includes a feature encoder and a text decoder. Therefore, during the training process, the training data is first subjected to text feature extraction to obtain... Then by using features The data is fed into a feature encoder, where high-dimensional feature information is obtained through a multi-head attention mechanism. Finally, high-dimensional information The text is fed into a text decoder to obtain the corresponding classification result. Then, the BCE loss function (binary cross-entropy loss function) is used to calculate whether the classification result is the same as the level 1 true label, thereby continuously learning and optimizing the encoder and decoder. The BCE loss function expression is as follows: ; In the formula, Represents the model's predicted values With real labels The smaller the value, the more accurate the prediction.

[0036] See Figure 2 ,like Figure 2 As shown, the corresponding feature vectors are then obtained by using a text feature encoder to process the level 1 ground truth labels. Multiply the high-order feature vector by the corresponding mixture weight value W_1, and then combine it with the first-level true label feature vector. Multiply by the corresponding weight W_G_1 and sum them to obtain the level 1 label mixed feature vector, which is then fed into the lower layer network for level 2 label classification learning, i.e.: ; In the formula, Let represent the input feature vector of the (i+1)th level. Let represent the input feature vector of the i-th level. This represents the mixing weights of the output features of the i-th model. This represents the mixed weights of the i-th level true label features.

[0037] A similar approach is used up to the 8th level of label classification.

[0038] It should be noted that the initial value of the mixture weight W_X for all high-dimensional features is 0.5, and the initial value of the mixture weight W_G_X for all true label features is 0.5. The trained multi-level context linkage learning model is obtained by training the single-level classification using the BCE loss function and by summarizing all levels of the overall network using the KL divergence loss.

[0039] The KL divergence loss expression is as follows: ; KL divergence is used to measure the difference between two probability distributions. and The smaller the difference value, the closer the distributions are.

[0040] In summary, the advertising classification method provided by this invention extracts dynamically associative classification feature vectors and combines them with hierarchical inference of a multi-level context-linked learning model. This allows for efficient handling of dynamically changing advertising categories without relying on predefined classification categories. By introducing a weighted feature fusion mechanism between levels during model training and inference, a tight logical association between multi-level classification labels is established, avoiding the loss of key contextual information and logical contradictions. Furthermore, by combining single-level precise learning with global learning, the model's classification capabilities are comprehensively improved, making it more robust in complex business scenarios with high requirements for classification accuracy and consistency. This invention achieves high efficiency and accuracy in advertising classification, effectively solving the problem of classification errors. It provides reliable data support for advertising performance attribution analysis, ensuring that subsequent adjustments to advertising delivery strategies align with actual needs, thereby improving the robustness of advertising delivery.

[0041] Based on the above method items, the present invention provides corresponding system items embodiments.

[0042] See Figure 3 , Figure 3 This is a structural block diagram of an advertising classification system provided in an embodiment of the present invention. The advertising classification system includes: The data acquisition module 21 is used to acquire advertising content from the platform side and category tag content from the advertiser side; Data processing module 22 is used to process the advertising content to obtain advertising description text and image understanding text, video understanding text or video subtitle text, and to concatenate the advertising description text with the image understanding text, video understanding text or video subtitle text to obtain advertising input text; The classification feature extraction module 23 is used to construct a multi-level classification system based on the classification label content, and extract classification feature vectors at each level based on the multi-level classification system using a text feature extraction model. The advertising classification module 24 is used to provide context information for subsequent classifications in an iterative manner based on the previous classification results through a pre-trained multi-level context linkage learning model, and to determine the classification of the advertising input text in the classification feature vectors of each level step by step to generate advertising classification results.

[0043] In one alternative embodiment, the data processing module 22 is configured to: The advertisement content is segmented and cleaned to obtain the advertisement description text; When the advertising content contains images or videos, the keyframes of the images or videos are understood through a visual language model, and the understanding results are segmented and cleaned to obtain image-understood text or video-understood text; the subtitles of the video image frames are extracted through a text recognition model, and the extracted subtitles are segmented and cleaned to obtain video subtitle text. The advertisement description text is concatenated with the image understanding text, video understanding text, or video subtitle text to obtain the advertisement input text.

[0044] In one alternative embodiment, the advertisement classification module 24 is configured to: The feature vector of the advertising input text is extracted by the text feature extraction module and used as the first-level input feature vector of the multi-level context linkage learning model. For each classification level, the input feature vector of the current level is matched with the feature vectors of each level to select the target classification for the current level; based on the input feature vector of the current level and the feature vector of the target classification, the input feature vector of the next level is generated. Continue until the last level of classification, and use the classification result of the last level as the advertising classification result.

[0045] It should be noted that the advertising classification system provided in this embodiment of the invention is used to execute all the process steps of the advertising classification method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0046] This invention also provides a terminal device, such as... Figure 4The diagram shown is a structural block diagram of a preferred embodiment of a terminal device provided by the present invention. The terminal device includes a processor 31, a memory 32, and a computer program stored in the memory 32 and configured to be executed by the processor 31. When the processor 31 executes the computer program, it implements the advertising classification method as described in any of the above embodiments.

[0047] In addition, embodiments of the present invention also provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the advertising classification method as described in any of the above embodiments.

[0048] When the processor 31 executes the computer program, it implements the steps in the above-described advertising classification method embodiments, for example... Figure 1 All steps of the advertising classification method shown. Alternatively, when the processor 31 executes the computer program, it implements the functions of each module in the above-described advertising classification system embodiment, for example... Figure 3 The functions of each module in the advertising classification system are shown.

[0049] Preferably, the computer program can be divided into one or more modules / units, which are stored in the memory 32 and executed by the processor 31 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0050] The processor 31 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor 31 can be any conventional processor. The processor 31 is the control center of the terminal device, connecting various parts of the terminal device through various interfaces and lines.

[0051] The memory 32 mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., and the data storage area can store related data, etc. In addition, the memory 32 can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, and a flash card, etc., or the memory 32 can also be other volatile solid-state storage devices.

[0052] It should be noted that the aforementioned terminal devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 4 The structural block diagram shown is merely a structural example of the terminal device described above and does not constitute a limitation on the structure of the terminal device. The terminal device may include more or fewer components than shown, or combine certain components, or use different components.

[0053] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An advertising classification method, characterized in that, include: Obtain advertising content from the platform and category tag content from the advertiser; The advertising content is processed to obtain advertising description text and image understanding text, video understanding text or video subtitle text. The advertising description text is then concatenated with the image understanding text, video understanding text or video subtitle text to obtain the advertising input text. A multi-level classification system is constructed based on the content of the classification labels, and the classification feature vectors of each level are extracted based on the multi-level classification system using a text feature extraction model. By using a pre-trained multi-level contextual learning model, contextual information is provided for subsequent classifications based on the previous classification results in an iterative manner. The classification of the input text for the advertisement is determined level by level in the classification feature vectors of each level, thereby generating the advertisement classification result.

2. The advertising classification method as described in claim 1, characterized in that, The pre-trained multi-level contextual learning model iteratively provides contextual information for subsequent classifications based on previous classification results, progressively determining the classification affiliation of the advertisement input text within the classification feature vectors at each level, and generating advertisement classification results, including: The feature vector of the advertising input text is extracted by the text feature extraction module and used as the first-level input feature vector of the multi-level context linkage learning model. For each classification level, the input feature vector of the current level is matched with the feature vectors of each level to select the target classification for the current level; based on the input feature vector of the current level and the feature vector of the target classification, the input feature vector of the next level is generated. Continue until the last level of classification, and use the classification result of the last level as the advertising classification result.

3. The advertising classification method as described in claim 1, characterized in that, The multi-level context-linked learning model is obtained through the following steps: Obtain advertising data, select a portion of training data from the advertising data to train a multi-level context linkage learning network, and obtain a primary multi-level context linkage learning model; The remaining advertising data is iteratively and hierarchically optimized using the aforementioned primary multi-level contextual learning model to construct an advertising training dataset. The advertising training dataset is used as the input to the primary multi-level context linkage learning model. The primary multi-level context linkage learning model is trained by hierarchical classification using a joint loss function to obtain a trained multi-level context linkage learning model.

4. The advertising classification method as described in claim 3, characterized in that, The joint loss function includes a single-level loss function for optimizing the classification accuracy at each level, and a global loss function for constraining the consistency of classification results across all levels.

5. The advertising classification method as described in claim 4, characterized in that, The step of using a joint loss function to perform hierarchical classification training on the primary multi-level context linkage learning model to obtain a trained multi-level context linkage learning model includes: The features output from the previous level classification are mixed with the current level's true label features according to preset weights to construct the next level training input features for the next level classification training. A single-level loss function is used to optimize the classification accuracy at each level, and a global loss function is used to constrain the consistency of classification results at all levels until all levels of classification training are completed, resulting in a trained multi-level context linkage learning model.

6. The advertising classification method as described in claim 1, characterized in that, The process of processing the advertisement content yields advertisement description text and image understanding text, video understanding text, or video subtitle text. The advertisement description text is then concatenated with the image understanding text, video understanding text, or video subtitle text to obtain the advertisement input text, including: The advertisement content is segmented and cleaned to obtain the advertisement description text; When the advertising content contains images or videos, the keyframes of the images or videos are understood through a visual language model, and the understanding results are segmented and cleaned to obtain image-understood text or video-understood text; the subtitles of the video image frames are extracted through a text recognition model, and the extracted subtitles are segmented and cleaned to obtain video subtitle text. The advertisement description text is concatenated with the image understanding text, video understanding text, or video subtitle text to obtain the advertisement input text.

7. The advertising classification method as described in claim 6, characterized in that, The visual language model is one of the following: a contrastive language-image pre-trained model, an image understanding model, or a guided language-image pre-trained model; the character recognition model is an optical character recognition model.

8. An advertising classification system, characterized in that, include: The data acquisition module is used to acquire advertising content from the platform side and category tag content from the advertiser side. The data processing module is used to process the advertising content to obtain advertising description text and image understanding text, video understanding text or video subtitle text, and to concatenate the advertising description text with the image understanding text, video understanding text or video subtitle text to obtain the advertising input text; The classification feature extraction module is used to construct a multi-level classification system based on the content of the classification labels, and extract classification feature vectors at each level based on the multi-level classification system using a text feature extraction model. The advertising classification module is used to provide contextual information for subsequent classifications based on the previous classification results through a pre-trained multi-level contextual linkage learning model. It determines the classification of the advertising input text in the classification feature vectors of each level step by step, and generates advertising classification results.

9. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the advertising classification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the advertising classification method as described in any one of claims 1 to 7.

Images